adrian_lab
Contact
THE A-TEAM · Your situation

Develop a working system with agents

This is a faithful English translation of the original Polish course.

In this chapter

AI-first here means that the agent receives needed knowledge and performs an agreed part of work while you check the result. It does not require rewriting the product. Start with a task already waiting to be done.

What changes in the way of working

Faster code writing alone will not remove waits between people. In one task, check whether you can go from need to demo with fewer handoffs. Open the history of a similar task and choose one handoff you will try to shorten.

Where work waitsHow to pass through it in the trial
Separate analysis handed onThe lead reads the task and documentation with an agent, returns to the expert with a concrete question and the missing knowledge.
Screen goes to one person, logic to anotherOne person leads a small feature slice from screen to data; another helps in an unfamiliar area and independently checks it.
Another contributor asks the expert the same questionAnswer, reason, and exception go into the feature description; the next person gets a link and an example.
A meeting is about reporting how much was doneOpen the working version, perform the task action, and decide a concrete observation.

Do not move all responsibility to one person without time and support. When an agent gets stuck, stop issuing more “fix it” attempts. Show the defect, expected result, and sources read to another person. First agree whether knowledge, access, or ability to make the change is missing.

Prepare one task

Use the system investigation and completed documentation. Point the agent to instructions, playbook, and sources needed for this task. Check whether it can explain a rule with evidence and recognize missing data. A domain expert confirms requirements.

Download · Text fileCard for a task led with an agentteam-zadanie-en.mdDownload

Attach the file to the conversation. If you have suitable fields in Jira or the plan, the agent should use them instead of creating a second work list.

Prompt for your agent
We want to take one current task with an agent: [link]. Read the attached file, playbook, and checked investigation of the area [source]. Show necessary context and questions that still have no answer. Propose a small plan: scope, lead, check, demo, and independent review. Use existing documents and the release process. Do not begin implementation until we agree the plan.

Run the trial with a second person

The implementer leads the change and the reviewer checks the result. Go through the demo and corrections, then tests and independent review. Also see whether another person or agent can take over the task using the written agreements.

Expand what actually helps

After the task, check quality, corrections, waiting time, and number of handoffs. Only then choose the next area. A rule change requires checking code and tests; a code change requires the description and, if used, the model to be current.

Result: a completed task and a concrete decision about what to preserve in the way of working. Documentation still has an owner. MiddleBrain describes joining domain knowledge to a code model; you do not have to build it for a first trial.