How a team works, three teams in detail, and ten more where you may recognise your own.
harnsy itself is built this way: by a team of AI agents with a lead.
An AI agent team in harnsy is the agents you already run (Claude Code, Codex, OpenCode), each given a role. You tell the lead what you want, and the lead gives each task to a role. The roles message each other, and the one who asked reviews the work. You answer only what needs you.
The same five steps in every team below.
Everything your team does is a task. The lead creates tasks and hands them out, the roles work them, and you accept the result.
Add CSV export to the reports page
Backend endpoint
Backend developerExport button
Frontend developerChecks
QA engineerThe hierarchy and one task from start to finish. The roles are a starting point, so add or drop any of them as the work needs.
Delete an empty workspaceabout 3 hours
Each role could run Claude Code, Codex or OpenCode: you choose per role.
Who does what, and one release from start to finish.
Site ready for a product release
Real · anonymized
An engineer runs it on their own project. Details of the building and the documents are left out.
You are an engineer who prepares design sections and needs the plan, the split of work and the checks in one place.
Not opened yet
TaskPlan the release of the water-supply and sewerage sections for a residential complex with high-rise buildings and an underground car park, in time for the state review, then start the work
The lead recorded the task and asked the engineer questions with buttons: the building type, the team size, which input data exist. It wrote a plan, and the engineer accepted it in the dashboard. At the engineer’s request the plan was split into 13 items, each for a role, with a checklist and a due date, and the lead opened six agents, each in its own terminal tab.
The lead first misread the building. The engineer’s answers to the specialists’ questions showed it; the lead sent a correction to all six, returned the delivered work for rework and rewrote the plan the same evening. The lead read every delivery and accepted it or returned it with a note, passed on the engineer’s decisions and gathered the specialists’ questions into one prompt with buttons.
Separately, the chief specialist wrote a research report with 25 sources on how water supply for very tall residential buildings is designed. Its conclusion changed a scheme chosen earlier: the lead gave the contradictions between standards to the standards checker and the change of concept to the engineer as a separate item. Work that needs the client’s data waits, blocked with the reason; the search for equipment suppliers went to the backlog. The team grew on the way: the lead added two roles, the chief specialist and the external networks engineer, when the engineer clarified the scope. The work is at the concept and calculation stage, and there are no final design documents yet.
The engineer saw the harnsy dashboard (the item tree with checklists, the attached plan, concepts, calculation and report, and the accept buttons), a terminal tab per agent, and button questions with short summaries in the lead’s chat. The team runs on Windows, all agents on Claude Code, in one project.
The engineer decides and signs. The team prepares the work; it does not give the conclusion.
Find the one closest to your work: who is on it, one task from start to finish, and the one question left for you. (Made-up examples; the water-supply team above is real.)
Any role could run any of the three.
You build a product with a few people and want builds, checks and notes to keep pace.
TaskAdd CSV export to the reports page
The lead splits the task: a backend endpoint, a frontend button, tests. The backend and frontend developers agree on the file shape by message. QA writes checks from the description. The reviewer builds the change in a clean copy and runs the tests. The release writer drafts the note, and the support liaison writes the answer for the first customer who asked. The lead accepts, then asks you one thing only you can answer: is export free or paid?
You answer questions with data and the path from a question to a chart is long and fragile.
TaskWhy did weekly signups fall 12%?
The lead gives the question to the analyst, which decides the decomposition itself: the funnel by source, by device and by release date. It asks the data engineer, by message, for a clean events table, and the data engineer delivers it with a description. The visualization developer builds three charts. The data-quality reviewer re-runs the queries and finds one wrong date filter, and the analyst fixes it. The lead accepts and asks you one thing: may the report name the partner whose traffic dropped?
You publish on a schedule in more than one language and want sources and facts checked before it goes out.
TaskWeekly newsletter, English and Spanish
The lead picks the topic from your list and gives it to the researcher. The researcher sends sources to the writer, who drafts. The fact-checker returns two unsupported claims; the writer removes one and finds a source for the other. The translator makes the Spanish version, and the layout preparer builds both. The lead accepts, and you approve the send.
TaskLaunch a spring campaign for a bakery chain
The lead hands the strategist the brief and lets the strategist split it: audience, channels, message. The strategist messages the media planner for channel costs, and the media planner replies with a table. The copywriter gets the three key messages from the strategist and drafts the ads. The reviewer checks the tone against the brand guide and returns two drafts with notes. The lead accepts, then asks you one question: which of two discount offers to use.
TaskClear Monday’s backlog of 140 tickets, from an export you give the lead
The specialists draft replies and submit them; the quality reviewer checks tone and policy and sends two back. The knowledge-base writer turns the three most repeated questions into articles. The lead accepts. Refund decisions above a limit wait for you, and you answer them in one batch.
TaskSurvey of methods for soil moisture mapping
The method reviewer compares the methods in a table. The data processor re-runs a published example. The writer drafts the survey, and the critic looks for claims without a source. You answer which three of the found methods the survey should focus on.
TaskPrepare the autumn catalog update: 300 products, from the catalog file you give the lead
The pricing analyst checks the new prices against the supplier list that the supplier liaison sends. The order-problem handler adds the top five complaints as product-page notes. The reviewer spot-checks 30 products. The lead accepts, and you decide the price change for two categories.
Three jobs where a person with the right to sign signs: a lawyer, an accountant, an engineer.
TaskReview a supplier contract before Friday
The clause extractor lists the clauses and messages the risk reader and the precedent finder. The redliner proposes changes, and the reviewer compares each proposed change with the extracted clause. A lawyer, you, makes the decision and signs. The team prepares the notes and the marks, never the conclusion.
TaskClose the month for three small clients
The document collector lists the missing receipts, the categorizer assigns accounts, and the reconciler compares them with the bank statement and flags five entries. The reviewer re-checks the totals. The accountant, you, approves and files. The team prepares entries and reports, never the filing.
TaskVentilation drawing set for a 600 m² workshop
The owner decomposes it: air volumes, duct sizes, the equipment list. The requirements engineer messages the normative limits, and the standards checker compares each number. The drawing-list preparer assembles the sheet list, and the reviewer flags two mismatches. The lead accepts, and the licensed engineer, you, checks and signs the set.
Agents work with what their CLI can reach: files, commands and the tools you connect to it. harnsy adds teamwork on top — messages, tasks, roles and shared documents — not integrations with other services.
Nothing here happens by itself. The lead adds a role when the work needs one, and you can see every change.
A product team does not start as it ends up. Over the weeks the lead adds roles when the work needs them and releases them when it is done: a portability audit, a terminal driver, an adapter for another harness. Some stay, the one-off ones leave, and every role that leaves writes its handover notes first.
The agents you already run, given roles. A lead sets the task and picks who does it; roles message each other, hand work over and check it; you set the direction and accept the result. harnsy connects them and has no model of its own.
Either. The lead splits it, or gives it to a role that splits it for the roles under it, as the analyst does in the data team.
No. You set the direction, answer the questions that need you, and accept or reject the result. harnsy shows you what waits for you.
Yes. Each role runs the harness you choose. Each role is an ordinary session of its harness, so its own subagents work inside it as usual.
Yes. Your agents use their own accounts with their providers; harnsy has no credits or tokens. The landing FAQ has the detail.
It stays on your machine, in ~/.harnsy. The landing FAQ lists exactly what harnsy sends and when.
No. They are examples. The water-supply design team is real, from one engineer, and anonymized.
In harnsy, one project runs one team of up to 10 agents on one machine. Several projects, more agents, or several machines working as one team take a harnsy Max key.
Three questions: your email, how many licenses (1–3), and what you would use it for. We read every application; the answer appears on a private page.
Ask for a harnsy Max keyOr start with one team: Install harnsy →
An application does not guarantee a key.