Session 11 — Project integration

Session 11 of 13PracticeDraft5 min read

Draft for the lecturer's approval. This is preliminary study material. It does not replace guidance published in Moodle, and it does not set any date or grade.

Goals

• Combine retrieval, a tool, a budget and a guardrail in the same run.
• Compare a single agent with two roles on the same questions.
• Write a README that another partner can run successfully.
• Record enough details in every run to reproduce it.

When the second role helps

The demo below is fixed: without a boundary, the two researchers bring two prices for the same supplier. A summarizer that picks the wording deletes a contradiction. A summarizer that keeps both sources at least does not make up a single price.

Two researchers and a summarizer. The messages come from fixed rules, so the same failure can be reproduced.

In lab 4 you replace the demo with code. If you cannot show a difference in the number of successes or in coordination failures, you do not write that the team improved the system.

Every hand-off between roles is a place where information can get lost. The summarizer sees only what the researchers wrote to it. If a researcher left out a source id, the summarizer cannot bring it back.

What to record in every run

To compare versions, and to find out why a given run failed, every run keeps:

• Which model and which version of the prompt.
• The package versions, or a lock file.
• The full log: every tool call, its arguments, its status and its result.
• The number of calls and the total time.

Without these details, "it worked yesterday" is not information anyone can check. A small prompt change is a new version, and you run the same cases on it again.

What another person checks

Following the README: installation, running the tests, and one question. If a file, a key or a command is missing, that is a finding of the session, not a minor glitch.

A README someone can run includes:

1. What the system does, in one sentence, and on which collection.
2. Install and run commands, from an empty folder.
3. How to run the set of cases, and where the results are saved.
4. Where to put a key, if one is needed, without it entering the repository.
5. What the system cannot do.

Swap repositories with another pair. Each pair tries to run the other's project using only the README, without asking questions. Write down every step where you got stuck. Then each pair fixes its own README according to the list it received.

Check your understanding

Two roles succeeded on the same number of cases as a single agent, and cost twice the calls. What do you write in the report?

Pick an answer to see the explanation. There is no grade here.

Reading

Anthropic, How we built our multi-agent research system, 2025. anthropic.com/engineering/multi-agent-research-system. Further reading: when several agents helped, and how much it cost.

Summary

What stays. A second role is justified only if it changes a measured number. Every run is recorded so it can be reproduced.

What breaks. A summarizer that deletes a contradiction. A README that only works on its author's computer.

Next: session 12, Guest lecture

The mark is saved only in this browser. It is not an academic record and is not sent to a server.