Multi-agent operating discipline Day 55 in production
한국어 See pricing
8 agents · running 55 days without a break

Two agents overwrite the same file.
git will not tell you.

You have two terminals open. The one on the left rewrites payment.ts. Three minutes later the one on the right pastes its own work onto the stale copy it read twenty minutes ago and saves over the same file. Nothing has been committed yet, so git raises no conflict. Both agents report back that they are done. Fifteen minutes of the left-hand agent's work is gone and nobody knows it. Add a third and a fourth agent and this stops being an occasional thing.

agentlock enforces exactly one rule. Declare the files you are about to touch, and keep off the files someone else has declared. Commit without declaring and a git hook stops you. Who claimed what, when, and who took it away is all on the record. The free edition is MIT open source. The paid edition puts deploys, approvals and completion reports under the same enforcement, and splits the work so cheap models generate while a stronger model reviews — which takes $200 a month down to $40.

Try the free edition first · MIT See what it would save you

The free edition is one file, no dependencies, and runs anywhere with Python 3.8 or newer. English and Korean are both supported. The figures below assume $200 a month in API spend, 80% of it going to generation, with generation moved onto a local model — every assumption is written out under the calculator.

This is not just my problem

The company that builds these tools documented it first

Anthropic states it plainly in the Claude Code documentation. They acknowledge the problem; the remedy they offer is for a person to split the work up and remember to keep it that way.

Claude Code · Agent Teams documentation

Official docs Remedy: manual discipline

“If two teammates edit the same file, you get overwrites. Split the work so each one owns different files.”

Splitting the work is the right answer. The problem is that a human has to remember it every single time. With three or four agents and sessions that run for hours, memory gives out — not if, but when. This tool takes that “just remember to” and moves it into code.

  • 12,974 file edits

    Spawning one teammate produced 151 instances editing the same codebase at once. claude-code #55586

  • One agent undoes another

    Reported to keep happening even after being told “Do NOT change any files you did not touch.” openai/codex #10681

  • Locks the task, not the file

    The same page says task claiming uses file locking to prevent races. Editing the file itself has no lock. Agent Teams docs

“Doesn't git worktree already solve this?”

For separate branches, worktree is right For one shared branch, git refuses

It does, and this tool is not a replacement for it. If you are giving each agent its own branch and running them in parallel, worktree is the correct answer. Run both together and they stay out of each other's way.

There is one place worktree structurally cannot reach. git will not let you check out the same branch twice. Branch state is shared between worktrees, and there is no way around it. Yet several agents sharing a single checkout is already the ordinary case. Claude Code's Agent Teams and subagents work that way, and so does a team where each person points their own agent at one repository. That gap is where this tool sits.

What it cannot do, stated plainly. Locking is per file, so two agents working on different functions inside the same file will still block each other. In a repository full of large files that means waiting in line for no good reason.

The locks are advisory. Skip the declaration, open the file in an editor and edit it directly, and nothing stops you. The git hook is the second line of defence, and that can be bypassed with --no-verify. What this stops is an honest mistake made by someone doing the work properly, not someone determined to get around it.

Semantic conflicts across different files are out of scope. Those belong to your tests and your reviewer.

And this is not an argument for running more agents in parallel. For small jobs, one at a time is faster. This exists to keep the people already running several out of trouble.

Working as a team

It doesn’t just stop overwrites. It makes them work as one team.

File locks are only the first button. For several agents to actually work together, more has to follow. Who owns what. Where a decision goes when one of them shouldn’t make it alone. Who picks up the work when one drops out. And how you know “it’s done” means anything. All of that is in here, as code rather than as a document nobody reads.

  • They divide the ground

    claim takes the files before the work starts. What someone else holds can’t be taken, and anything outside your area is refused at commit time. Locks expire, so an agent that dies doesn’t wedge the repository.

  • Decisions go upward

    msg and inbox let agents talk. Instead of deciding alone, the hands stop and say “this needs a decision” — and the head picks it up and answers.

  • Work survives a dropout

    handoff leaves a note behind. When a session dies or the context window runs out, the next agent reads what was in flight and continues instead of starting over.

  • “Done” has to be earned

    report refuses to file when the “could not verify” section is empty. A blank there isn’t a lie, but it reads like one to whoever comes next. That misreading gets closed off structurally.

  • One deployer, with approval

    request and approve put a human in front of deploys, schema changes and anything you mark risky. Writing it in the ledger is not the same as being approved.

  • Records that can’t be quietly edited

    ledger is append-only and hash-chained. Change or delete a single line and it tells you exactly which one was tampered with. The whole team reads the same history.

And that is what makes cheap models usable

discipline first the saving is downstream

The real reason people don’t put a local or low-cost model on serious work isn’t capability. It’s that you can’t trust the output when it works unsupervised. It edits files that belong to someone else, reports work as verified when it wasn’t, deploys without asking, and leaves no record of what it did. Then a human has to re-check everything, and at that point you may as well have paid for the expensive model from the start.

Put the six things above in code and that arithmetic flips. The bulk generation goes to the cheap side; only judgement and review stay on the strong model. When the cheap side steps over a line the tool stops it, decisions it shouldn’t make travel upward, and what it failed to verify is written into the report. $200 a month becoming $40 is the result of that — not because the model is cheap, but because the cheap model became something you can rely on.

The free agentlock implements the first of these — file ownership. It is MIT licensed and fine to use on its own.

The other five — crew talk, handoff, verified completion reports, approval gates and the hash-chained ledger — plus per-role identity prompts and the cost-routing setup, are in the paid edition.

Cost structure

What costs money is volume, not judgement

Almost all of the bill comes from output tokens. Writing new code produces a lot of output; reading it and pointing out what is wrong produces very little. The same task lands differently on the invoice depending on who does it.

  • Generation is long output

    Writing one file from scratch costs thousands of tokens. So does refactoring, so does writing tests. This is where the money goes.

  • Review is short output

    Reading code and answering "this part is wrong" is a few lines of output. The input may be long, but input is cheap next to output pricing.

  • So splitting them brings it down

    Send the volume work to the cheap side and keep only the judgement upstairs. Move it onto local hardware and the token cost on that side goes to zero.

Price per million output tokens

Bar length is the price. Same volume generated, different amount billed.

  1. Frontier model $25
  2. Cheap remote model $5
  3. Local model $0

Per million output tokens. The gap between a frontier model and a cheap remote one is five to one. Local means zero token cost — electricity and the hardware itself are not in that number. Input tokens are priced separately and are usually cheaper.

Cost calculator

Put your own numbers in

It updates as you move the controls. Nothing is stored and nothing is sent to a server. The arithmetic happens inside this page.

Inputs

US dollars · type an exact figure if you prefer
It varies by team. If you are not sure, leave it at the default 80%.

Where generation moves to

Result

New monthly cost
$40
Saved per month
$160
Saved per year
$1,920
Subscription pays back in
~5days
New monthly cost $40 Cut from the bill $160

How this is calculated · assumptions

  • The share you set as generation moves to the cheaper side. The rest — review, design, judgement — stays on the frontier model.
  • Local is counted as $0 in token cost. Electricity and hardware purchase are not included.
  • Cheap remote is figured at one fifth: $25 against $5 per million output tokens.
  • Payback is the Solo subscription of $29 ÷ (monthly saving ÷ 30 days). It assumes the daily saving holds steady.
  • At the default 80% the bill drops to a fifth. A team at 90% generation moving all of it local gets to a tenth. A lower share means a proportionally smaller cut.
  • Weaker generation means rework. That time is not in this calculation.

It is not only about saving

Work you used to turn down starts to add up

When the cost floor drops, you do not just keep more of the money. The range of jobs you can quote on gets wider.

What is left of a $600 job once the AI bill comes out

AI $200 → $400 left AI $40 → $560 left Difference $160 · about five developer-hours

Say you take a $600 job. If the AI costs $200 you are left with $400, and once you price in your own hours there is not much reason to accept it. The same job with $40 of AI leaves $560. That $160 difference is roughly five hours of developer time at a $30-an-hour blended rate. This is not an argument for cutting your rate to win more work. It is that jobs you used to turn down because the arithmetic did not work now do.

So why isn't everyone doing this already

Because a cheap model left alone gives you no way to know what it did

Everything below actually happened while running several agents for 55 days. The more work you hand to the cheap side, the more often it happens.

  • Agent A fixed a file. Agent B put the whole thing back to an older version.
  • Two agents changed the same function in two different ways, and both reported it done.
  • Something that worked yesterday is broken today, and there is no record of who touched what, or when.
  • Every new window the agent behaves like a different person, so you explain the whole thing from scratch again.
  • One agent deployed, another deployed on top of it, and the rollback order came apart.
  • An agent goes quiet mid-task and someone has to pick it up, with no way to tell how far it got.

The cost here is not the lost code. It is the time a person spends re-checking "is the current state the correct state?" over and over. So most people give up on the cheap side and push everything through the expensive model — paying more rather than trusting something they cannot verify. The nine disciplines below aim at exactly that point. Once what was done, what was skipped and what was touched are all on the record, the cheap side becomes safe to give work to.

Why it happens

Not because the models are bad, but because three things are missing

Each agent only sees its own task. Give it no way to see what the one beside it is doing and the more diligent the agent, the more it overwrites.

  • No ownership

    Every agent can write every file. With no boundaries, a collision is not an accident. It is a matter of time.

  • No ordering

    Work starts without being announced. Two agents discover they are in the same place only once they are both already in it.

  • No memory

    When the conversation window ends, the context goes with it. The next agent sees only the code that is left and guesses at the rest.

Thin code, thick discipline

4,250 lines of operations code 90-line sync server $0 a month to run

Try to solve this with a large framework and you have bought yourself a second job: maintaining the framework. What actually held up for 55 days was not code, it was discipline. The code was kept to the minimum needed to enforce it. So what is for sale here is not code. It is nine disciplines that survived running a real service. With those standing, the output of a cheap model becomes usable — and only then does the arithmetic above hold.

Track record

Records instead of claims

The numbers below accumulated inside a live commercial service in production. This is not an environment built for a demo.

  • Uninterrupted operation

    55days

    20 Jun 2026 to now, still running

  • Agents running at once

    8

    Roles fixed and kept separate

  • Messages between agents

    1,990

    About 36 a day

  • Automated meeting records

    960

    Unattended standing meetings · about 17 a day

  • Handover documents

    32

    Written every time ownership changes

  • Total operations code

    4,250lines

    Plus a 90-line cloud sync server

  • Monthly running cost

    $0

    No paid infrastructure behind it

Days in operation

One tick is one day. Not a single day missing between the start and today.
Started 20 Jun 2026Today · running

A day it stopped would show up as a gap here. There are no gaps yet.

The company and the service are not named. It is a live service, and naming it would mean disclosing its internals along with it. The numbers and the disciplines are published; the identity is not. If that trade does not sit right with you, do not buy.

The nine disciplines

Nine mechanisms that make a cheap model safe to put on real work

What you need in order to hand work to the cheap side is not a smarter model. It is a mechanism that forces a record of what was done. Leave the unverified field empty and the completion report is rejected. Touch a file outside your assigned area and the commit is blocked. Anything hard to undo does not move without a human approving it. All nine titles are published right here, because there is no reason to hide them. The hard part was never knowing the titles. It is the procedure that makes them hold every day.

  1. DISCIPLINE 01

    Fixed role separation

    Every agent gets a fixed area and a fixed tier. The hands that produce volume and the head that makes the calls are never the same agent.

  2. DISCIPLINE 02

    File ownership plus interface contracts

    Split which files each side touches and connect them only at agreed seams. Editing across a boundary becomes a change to the contract.

  3. DISCIPLINE 03

    Mandatory start-of-work declaration

    State what you are about to do before you touch code. This is the one discipline the free tool, agentlock, already implements in code.

  4. DISCIPLINE 04

    One deployer, no exceptions

    Only one agent deploys. Ban cross-deploys and the rollback order stays intact.

  5. DISCIPLINE 05

    Succession protocol when someone goes dark

    Define what to read, and in what order, to take over when the other side stops answering. Handled as procedure, not judgement.

  6. DISCIPLINE 06

    Identity-pinning prompts

    Pin a self-image to each tier. The hands are told not to decide alone; the head is told not to do volume work itself.

  7. DISCIPLINE 07

    Completion reports must carry evidence

    Write down what was verified and what was not. Leave the unverified field empty and the completion report itself is refused.

  8. DISCIPLINE 08

    Approval gates

    Deploys, money and database schema changes go through a human. This is where the cheap side is kept out of the rooms that are hard to walk back from.

  9. DISCIPLINE 09

    Append-only ledger

    Work done is only ever appended to a hash-linked record. Change or delete a single line and it tells you which line was tampered with.

Without these nine standing, you cannot hand work to the cheap side. Output where nobody knows what was done, what was skipped or what was touched has to be reviewed from scratch by a person, and that time costs more than the money saved. The written document for each discipline, the identity-prompt templates for each agent, the handover form and the conflict-response procedure all ship with the pro edition as a document kit. There is nothing extra to buy. The version where a tool enforces the same nine instead of a person remembering them is agentlock pro.

Role-split setup

The split between hands and head ships as commands

These went in recently, and only things that actually work are listed. The split points are already built, so you do not have to design the structure yourself.

  • setup --preset 4

    Split and set it all up in one pass

    Splits roles into hands (local) and head (frontier model) in a single command. Assigned areas, the designated deployer, approval points, git hooks and the per-tier identity prompts all come out of this one command.

  • env

    Check the local runtime

    Detects whether Ollama, LM Studio, llama.cpp or vLLM is up. Reads your machine's memory, tells you what model size it can carry, and warns you when a local role is assigned but no runtime is running.

  • cost

    Work out the saving

    Enter what you spend now and your generation share, and it returns the new monthly cost, the monthly saving and the annual saving. Same formula as the calculator further up this page.

  • route

    Who takes which job

    Give it a kind of task and it tells you whether the hands or the head take it. The point is to fix the criteria once so a person is not deciding it every time.

  • msg / inbox

    The channel between agents

    How the cheap side escalates the calls it should not be making alone. What was escalated and what came back both stay on the record, so you can retrace later why a decision went the way it did.

  • Generated identity prompts

    The wording changes with the tier

    The hands get "do not decide on your own." The head gets "do not do volume work yourself." It keeps the cheap side escalating where it should, and keeps the expensive side from burning money on output.

What the hands take

Local or cheap remote · the long-output side

  • Implementation
  • Fixes
  • Refactoring
  • Writing tests

What the head takes

Frontier model · short output, heavy judgement

  • Review
  • Design
  • Deploys
  • Schema changes

The split criteria live in route and can be changed to suit your team. If you plan to use the local side, run env first and check that your hardware can carry it.

Pricing

Start free, and move up only as far as you need to

It is a subscription. Pay and the licence key and files are issued on the spot — there is no human approval step to wait on. Annual billing is 20% cheaper, and you can cancel yourself from the customer portal at any time. The payback figures below assume a saving of $160 a month ($200 spend, 80% generation share, moved local). For your own numbers, use the calculator.

agentlock Free · open source

A CLI that puts discipline 3 — the mandatory start-of-work declaration — into code. It makes an agent declare what it is about to do before touching a file, and blocks it when that overlaps work already claimed. Install it, run it for a single day, and you can see where your collisions are coming from. agentlock pro below is the version that enforces the other eight as well.

Get it on GitHub

Solo

One developer · up to 5 machines

$29 / mo

$278 a year — two months off. At $160 saved a month it pays for itself in 5 days.

This is not a document you read and try to follow. Break a rule and the commit or the deploy stops right there.

  • The free edition puts 1 of the 9 disciplines into code. The pro edition puts all 9 into code
  • One setup --preset 4 splits hands from head — assigned areas, designated deployer, git hook installation and per-tier identity prompts are all generated for you
  • env checks your local runtime and machine memory, cost works out the saving, route shows who takes which job
  • Committing files from someone else's area, completion reports with no evidence, and deploys without approval are all blocked on the spot
  • The record is hash-linked, so changing or deleting even one line tells you exactly which line was tampered with
  • The full discipline document kit is included. Nothing extra to buy
  • No dependencies · Python 3.8 or newer · a single file
  • Scope of use: the one person who subscribes. On any project you work on yourself — personal work, freelance client work, work at the company that employs you — there is no restriction
  • Committing it into a shared team repository, or passing the files to someone else, needs Team or above
Start a subscription
Recommended

Team

One repository · up to 10 developers · 20 machines

$79 / mo

$758 a year — two months off. Counting one person's saving alone ($160 a month), it pays back in 15 days.

This is the tier if several of you share one repository. Who did what stops being a memory and becomes a record.

  • Everything in Solo
  • Up to 10 developers on one repository
  • A shared hash-linked ledger, so completion reports can be verified instead of believed
  • Approval gates — one fixed deployer, with approval enforced on deploys, schema changes and anything marked high-risk
  • Handover and inbox — someone else can pick up the work of a teammate who is away
  • Identity prompts that hold each agent to the role it was given
  • Email support
Start a subscription

Business

No limit on repositories, people or machines

$199 / mo

$1,910 a year — two months off. On one person's saving alone that is 37 days; across a team it is far shorter.

Once you have several repositories, or you find yourself counting heads, this tier is the cheaper one.

  • Everything in Team
  • No cap on developers, no cap on repositories
  • Priority support — you go to the front of the queue
  • Rollout design help — we work out with you which agent takes what, where the approval gates belong, and how work is divided between the cheap and the frontier model
Start a subscription

All three tiers are the same file. Only the scope of use differs

20% off annual Change plan any time Cancel it yourself

Solo, Team and Business ship exactly the same file, with no difference in features. What changes is how many people may use it and across how many repositories. So start small, and when the team grows, move the plan up in the customer portal. Nothing to download again, nothing to reinstall. Cancelling is in the same place and you never have to contact us to do it.

Payback is the price divided by the daily saving (monthly saving ÷ 30 days). For Team and Business the figures are deliberately conservative — they count one person's saving only, and shorten as more people use it. If your saving is smaller than the calculator's default, payback takes proportionally longer.

The price shown is the price charged. VAT or sales tax is calculated for your country at checkout and shown alongside it. The receipt is emailed automatically the moment payment goes through.

The licence key and files are issued the instant you pay. Nothing to wait for. If you change machines, you can release the old device registration yourself in the customer portal.

English and Korean are both supported. Choose the interface language with AGENTLOCK_LANG; leave it unset and it follows the language of your environment. The documentation ships in English and Korean too.

You do not need to ask before buying — just pay. If you are unsure which tier fits, describe your situation instead. Team size and how you use the repository is all we need to point you at the right one.

Frequently asked

The things you are probably about to ask

Can a local model actually write decent code?

Not as well as a frontier model. Let me be direct about that. Given the same requirements it misses more of the conditions, and it touches the wrong file more often.

That is exactly why review is attached. The structure only works when the head reads what the hands wrote and sends it back. Running local with no review is not something I would recommend — you will spend more time later finding the cause than you saved.

Put another way, what this sells is not "cheap models are good." It is "the output of a cheap model becomes checkable." Leave the unverified field empty and the completion report is rejected; files outside your area cannot be committed; what was done stays in a hash-linked record.

What kind of machine do I need?

Memory is the criterion. 16GB is tight — small models will run, but the range of work you can hand over is narrow. From 32GB it becomes genuinely useful, and at 64GB or more you can move most generation work local.

Run the env command to see where your machine sits. It detects which of Ollama, LM Studio, llama.cpp or vLLM is up, and tells you from your memory what model size you can carry.

If the hardware is not there, use a cheap remote model instead of a local one. The saving is smaller, but there is no hardware requirement at all.

Can I use a cheap remote model instead of a local one?

Yes, and the structure is identical. Generation moves down and review stays up. Whether the bottom is local or a cheap remote API, the disciplines apply the same way.

It is not free, though. Per million output tokens it is $25 against $5 — a fifth. On $200 a month with an 80% generation share, local comes to $40 a month and cheap remote to $72. Choose "cheap remote model" in the calculator to run it on your own figures.

Does it really come down that much?

It depends on your generation share. The calculator assumes the following: the share you set as generation moves to the cheaper side while the rest stays on the frontier model; local costs $0 in tokens; cheap remote is one fifth; and payback is the price divided by a daily saving that is assumed to hold steady.

At the default 80% the bill drops to a fifth. A team at 90% generation moving all of it local gets to a tenth. If you are already mostly reviewing and designing, and your generation share is 30%, the cut is proportionally smaller.

What is not in the calculation, stated plainly: electricity, hardware purchase, and the time spent reworking output when generation quality falls short. Which is why I am not asking you to believe the numbers. Put your own into the calculator.

Can I get a refund?

Before access is issued, a full refund, always.

Once you have it, tell us within 14 days of delivery and we refund without asking why. Write to zenesisgp500@gmail.com — English enquiries are handled by email. There is also a KakaoTalk open chat, but it is a Korean-language chat. Customers in South Korea additionally have statutory withdrawal rights, set out under Cancellation and refunds.

If you simply picked the wrong tier, moving up and paying the difference is usually a better outcome than a refund. Either way, we do whichever you prefer.

Our setup is a bit different. Does this still apply?

The disciplines are about boundaries and ordering, not about any particular tool's name. Whatever agent you run, "who writes which file", "where work gets declared" and "who deploys" are all still needed.

Repository structure does vary, though. The kit's forms and procedures were written on the assumption that you will adapt them to your structure, and the pro edition has you name the assigned areas, the deployer and the approval points during setup. Different structure, same setup step — you just point it at yours.

Free edition or pro — which one do I need?

Run the free edition for a day and then decide. The free edition puts 1 of the 9 disciplines — the start-of-work declaration — into code. The other 8 are things a person has to remember.

The pro edition puts all 9 into code. Files in someone else's area can neither be locked nor committed; leave the "could not verify" field of a completion report empty and the command is refused; a deploy runs only when the one designated person holds a human approval, and that approval is spent once it is used. Nobody has to check whether the rules are being followed.

Draw the line roughly here. One person with one agent: the free edition is enough. Two or more agents at once, or sharing a repository with human colleagues: pro.

Which pro tier to buy is set out under “Which tier should I buy?” below.

How is the pro edition licensed?

It is a subscription with three tiers. All three ship exactly the same file, with no difference in features. What changes is how many people use it and across how many repositories.

Solo · $29/mo ($278/yr) — one developer. On any project you work on yourself — personal work, freelance client work, work at your employer — there is no restriction. Up to 5 machines can be registered. Committing it into a shared team repository, or handing the files to someone else, is not covered at this tier.

Team · $79/mo ($758/yr) — up to 10 developers on one repository, with up to 20 machines registered. Commit it into your internal repository and share it with the team, or adapt it to your environment.

Business · $199/mo ($1,910/yr) — no limit on people or repositories, and unlimited device registrations. Priority support and rollout design help are included.

None of the three tiers permits resale, public posting or redistribution to third parties. Taking the files outside your organisation, or bundling them into another product you sell, is not allowed at any tier.

The licence key is valid for as long as the subscription is. Cancel and the key is revoked, and under the agreement you delete the copies you hold.

Which tier should I buy?

A one-person company or a freelancer working alone: Solo. It is your own working environment, and using it on projects you have taken on is unrestricted. Once you commit the files into a client's or a team's shared repository for several people to use, that is Team territory.

Human colleagues sharing one repository with you: Team. Up to 10 people. Spread across ten, that is under $8 per person a month.

Several repositories, or more than 10 people: Business works out cheaper, and you stop counting altogether.

If you cannot tell right now, start on Solo. You can move up from the customer portal at any time, and you are only charged for the period you used.

Can I move up to Team or Business later?

Yes. Change the plan in the customer portal. You do not need to contact us.

The file is the one you already have. Nothing to download again, nothing to reinstall. Only the licence key changes tier.

Billing is prorated for the remaining period, so what you have already paid is not thrown away. Moving back down works the same way, in the same place.

Is this any use if I only run one agent?

About half of it is. The collision disciplines (2, 4 and 8) get less exercise, but start-of-work declarations, identity pinning, verification evidence and handovers all work just as well on your own. The difference shows most when a working session gets cut off and you pick it up again.

And if you share a repository with human colleagues, the rest starts earning its keep soon enough.

How do I pay?

By card. Press the button under How to buy below and the checkout opens. All six plans sit on one screen, so you can click between them and compare.

There is no need to contact us first. Your email and card details are all it takes. VAT or sales tax is calculated for your country and shown with the total.

Payment and receipts are handled by Polar Software, Inc. as the reseller. Your card statement will show the Polar name.

How do I receive the files?

Straight from the screen the moment payment completes. The licence key and the file download appear together, and the same thing goes out by email.

You can return to the customer portal at any time and download again. Changing machines or losing the file is not a problem.

When you move to a new machine you can release the old device registration yourself in the portal. No need to contact us.

Do I get documentation I can give to accounting?

A receipt is emailed automatically the moment you pay. It is issued in the name of the reseller, Polar Software, Inc., and you can retrieve it from the customer portal at any time.

Tick “I’m purchasing as a business” at checkout and you can enter your company details and tax registration number, both of which then appear on the receipt.

It is not, however, a Korean statutory tax invoice or cash receipt. If you specifically need those documents, tell us before you buy.

Do I get the actual source?

The free agentlock is published as open source, so take it as it is. agentlock pro is a single Python file, handed over as it is. Every line is visible and you can read it yourself if you want to.

The full 4,250 lines of operations code is not for sale, though. That code grew around my environment, and handing it over would not make it run in another team's repository. What has to move is not the code, it is the disciplines — and the part that moves them is what was cut out into the pro edition.

Do I get updates after I buy?

Every revision released while you are subscribed is included, at no extra cost. When a new version lands, take it from the customer portal.

Both the discipline documents and the pro edition keep getting corrected as the system keeps running. I do not promise a release cadence, though.

How long do I have to wait?

You do not. The licence key and files are issued the moment payment completes. There is no human approval step in the way.

Plan changes and cancellations are also immediate in the customer portal. Only enquiries are answered by a person — usually within a few hours, and at worst within one business day. If it is urgent, the KakaoTalk open chat is faster, though it is a Korean-language chat; for English, email is the way.

What happens if I cancel?

One button in the customer portal. No need to contact us, and no reason required.

Cancelling leaves you everything you have already paid for, until that period ends. After that the licence key is revoked and you are not billed again.

On cancellation the licence agreement requires you to delete the copies you hold. The key is revoked and the pro features stop. The free agentlock is open source, so that keeps working.

Within 14 days of payment, ask for a refund rather than just cancelling and you get the whole amount back — see Cancellation and refunds.

Is the tool itself in English?

It is. Both the program interface and the documentation are available in English.

Set AGENTLOCK_LANG=en to force it. Leave it unset and it follows the language of your environment. Different people on the same team can run it in different languages; the records that accumulate in the repository use the same format regardless, so nothing gets mixed up.

The manual, the quick start, the licence guide and the whole document kit ship in English and Korean. An English reference copy of the licence agreement is included as well. Note, though, that the Korean text of the agreement is the governing version and the English is provided to aid understanding.

If you find wording that reads badly, tell us. The translation table sits in the source, so you can rewrite it in the phrasing your team uses and the program still runs.

How to buy

Pay, and it is issued on the spot

There is no need to contact us first. Press the button below, pick a plan at checkout, and the licence key and files come out as soon as payment completes. There is nobody to wait on.

  1. Pick a plan at checkout and pay

    One link holds all six: Solo, Team and Business, each monthly and annual. Click between them on the screen to compare. Annual takes 20% off. Pay by card, and VAT or sales tax is calculated for your country and shown with the total.

    Payment and receipts are handled by Polar Software, Inc. They act as merchant of record, so your card statement shows the Polar name. The receipt is emailed automatically as soon as payment goes through.

  2. The licence key and files appear immediately

    When payment completes, a licence key in the form AGENTLOCK-XXXX-XXXX-XXXX and the file download appear on that same screen. The same thing goes out by email. You can return to the customer portal at any time to download again, or to release a device registration after changing machines.

    Registered machines: 5 on Solo, 20 on Team, unlimited on Business. Those numbers allow for replacing a laptop or reinstalling.

If the tier is unclear, ask before you pay. Team size and how you use the repository is all we need to tell you which one fits. Enquiries are usually answered within a few hours, and at worst within one business day. If a day passes with no reply, check your spam folder and then send it again.

Buy now

All six plans on one screen. Click through, compare, choose.

Open checkout

What you can choose at checkout

  • Solo $29/mo · $278/yr
  • Team $79/mo · $758/yr
  • Business $199/mo · $1,910/yr

Choosing annual takes two months off. Switching between monthly and annual later is also done in the customer portal.

Email
zenesisgp500@gmail.com
Open chat
KakaoTalk open chat — Korean-language chat

You can buy without contacting us at all. Enter your email at checkout and the key and files go to that address. For enquiries in English, use the email above.

How delivery works, stated plainly. Files are issued as a download tied to the subscriber's account, and the licence key is valid only while the subscription is. Cancel and the key is revoked, the pro features stop, and under the licence agreement you delete the copies you hold. The free agentlock is open source and keeps working. You can cancel from the customer portal yourself, without contacting us. What happens to the remainder of a paid period, and how refunds are handled, is set out just below under Cancellation and refunds.

Cancellation and refunds

Cancel it yourself. 14-day refund, no questions asked

Because this is a subscription, two separate things exist. One is cancelling — not being billed again. The other is a refund — getting back what you already paid. Written the way the process actually runs, not copied out of a statute.

  • Cancelling · any time, yourself

    Cancel with one button in the customer portal. No need to contact us, and no reason to give. Cancelling leaves you everything you have already paid for, until that period ends. After that the licence key is revoked and you are not billed again.

  • Refund · 14 days, no questions

    Tell us within 14 days of payment and we refund the full amount without asking why. No form to fill in, no explanation needed. Email or the open chat, whichever suits you. It takes effect the moment you send it.

  • Money back · within 3 business days

    It is a digital file, so there is nothing to send back. We only revoke the licence key. The full amount you paid is returned within 3 business days of us receiving your request. A card reversal can take a few days longer on the issuer's side.

It renews automatically, and we warn you before it does

Monthly · same date each month Annual · 7 days' notice Cancel any time

The subscription renews automatically until you cancel, charging the same amount to the payment method on file. Monthly renews on the same date each month, annual on the same date each year. For annual subscriptions we email you 7 days before renewal. Cancel then and the following year is not charged. We also let you know in advance when a card is close to expiring, so nothing gets cut off by a failed payment.

For customers in South Korea, the following statutory rights also apply

Withdrawal within 7 days 3 months / 30 days if it differs from what was advertised Act on Consumer Protection in Electronic Commerce

If you are buying from South Korea, the Act on Consumer Protection in Electronic Commerce gives you a statutory right of withdrawal. Within 7 days of the date of payment you may withdraw and be refunded in full. No prescribed form and no explanation are required — telling us by email or open chat that you wish to withdraw is enough, and it takes effect at the moment you send it.

If what you received differs from the description and advertising on this page, or the contract was not performed as agreed, the period is longer. You may withdraw within 3 months of the date of supply, and within 30 days of the day you became aware of it or could reasonably have become aware of it. The way you notify us and the refund within 3 business days are the same as above.

The 14-day no-questions refund above is offered voluntarily and is longer than the statutory 7 days. It sits on top of these rights and neither replaces nor reduces them.

Where to send a refund or withdrawal request — zenesisgp500@gmail.com, or the KakaoTalk open chat (a Korean-language chat). Either reaches us; for English, use the email.

Whichever route you use — the 14-day policy or the statutory right — the process is the same. Send us a note, we revoke the key, and the money goes back within 3 business days.

Payment, receipts and refund processing all run through the reseller, Polar Software, Inc. Send your request to the address above and we handle it from there.

The seller's name, address and contact details are at the very bottom of this page, along with how personal data is handled.