The team has changed. Code is a tool, the product is the job.
Agents now write the code, run the tests and prepare the release. What a team needs are product engineers: people for whom code, testing and deployment are tools for making a product, and who answer for the result.
agentReads the ticket and the code, asks two questions
personAnswers and approves the plan
agentWrites the code and the tests in its own copy
agentOpens a merge request, the pipeline is green
personReads the change, presses “To production”
agentRolls the release out product by product
Agents do the typing. A person makes every decision.
What changed
From roles by tool to one owner of the result
Before
A task goes from analyst to developer to tester to DevOps and waits in a queue at every hand-off.
Each person answers for a tool: the code, the tests, the pipeline. Nobody answers for whether the user got what they needed.
Most of the time goes to typing: code, test cases, release notes.
Now
One product engineer takes the task from the question to production; agents do the typing.
The person sets the goal, checks the plan, reads the change and decides when it ships.
Time goes to what agents cannot do: understanding the user, choosing what to build, judging the result.
Product engineer
What the job is now
Understands the user and the businessKnows why the task exists and how to tell that it worked.
Sets the task for an agentThe goal, the limits, examples and the check that proves it is done.
Reads what the agent wroteReviews the change as a senior colleague’s work: logic, data, security.
Answers for productionShips, watches the numbers and rolls back when needed.
What we built
Autonomous agents in a working team
In the development team we lead, agents are part of every working day. Three things we built carry that work.
ai-cli
The agent takes a task from the board
A server that runs Claude Code, Codex and Cursor on the whole codebase, with read-only access to logs and data. The team lead presses a button on the task board: the agent reads the ticket, writes the code in its own copy, opens a merge request and links it to the task. The next button merges it once the pipeline is green and rolls the release out product by product. A conflict with the main branch the agent resolves too.
Do the task
To production
Rebase
AI-Research
Questions before code
Before anyone writes code, an agent reads the ticket and the code it touches and asks the author what is missing, right in the task. When the answers are in, the task is marked ready. Developers stop finding gaps halfway through the work.
AI-clarification
AI-ready
CI
The codebase as the agents’ workplace
The main repository carries skills and rules for agents across the whole cycle. The same skills run on a laptop, in the team chat and on the server, so an agent works the same way wherever it is started.
task analysis
implementation
backend and frontend QA
merge request review
release
on duty
production incidents
Tools
We use every agent there is
We are not tied to one vendor. Each task goes to the agent that handles it best today, and a new model is tried on real work, not on a demo.
The rules and skills live in the repository, not in a tool. Switching the agent does not mean starting over: in ai-cli it is the model name in the request.
Agents
Claude Code
Codex
Cursor
Models
Anthropicall
OpenAIall
Grok
Qwen
Kimi
+ everything in Cursor
For your team
What we can set up
Agents on your codebaseRules, skills and safe read-only access to logs and data, so agents work on your code the way your team does.
One process, end to endWe pick what eats the most time: tasks from the board, QA, releases, on-duty checks. We build an agent that does it and the point where a person signs it off.
A pilot on real tasksWe work on your own backlog and measure the result in your numbers, not ours.
Product engineersWe teach the team to set tasks for agents, review their work and own the result in production.
The point
AI is not a cure-all that brings more customers. It is a tool for taking more of the market.
Customers come for a product that solves their problem, and a chatbot on the website does not change that. What AI changes is how much a team gets done: more ideas tested, releases sooner, more segments, languages and stores covered. That reach is where the extra share comes from.