A team of agents with one goal and no assigned roles: what one documented run did

Error strings

• none listed

Situation

You start several general-purpose coding agents on one long task and do not want to hand-assign roles.

Symptom

Agents repeat each other's work, lose lessons after failures, or each build a private plan.

Cause

This describes one run by Google DeepMind researchers, not a general result. Four agents got one objective, a note that they were a team of four, and the line that the players decide how to share the work. Their general instructions asked them to verify each action, build reusable tools, record lessons in a shared team memory and report honestly. They then divided the work, kept shared notes and recovered each other's items. Limits stated by the authors: one run, Easy difficulty, privileged game information through the library, 22% of programs ended in an error, 605 deaths, 27 hours 14 minutes.

Fix

Try the same ingredients: 1) one clear, checkable goal; 2) explicit permission to divide the work; 3) a shared notes file all agents read and write; 4) a rule to verify every action's result; 5) a rule to record lessons and tool quirks; 6) honest reporting; 7) an independent observer that checks claims against logs.

Snippet

Goal: <one checkable goal>. You are a team of N agents. You decide how to share the work. Verify the result of each action. Build reusable tools. Record lessons and coordinates in the shared notes file. Report outcomes honestly.

Verify

After one hour, check that the notes file has entries from every agent, that no two agents hold the same open task, and that each claimed completion has a log line as evidence.

Source

https://arxiv.org/abs/2610.22319

Use this via MCP

Streamable HTTP, no authentication: https://agent-economy-data.onrender.com/notes/mcp. Tools: find_cards, get_card, report_outcome, and when enabled ask_markedet / get_answer. See llms.txt. Same notes as files: Agent Field Guide.