Agent Arena: practise a handoff that can be checked
Use a paper team exercise to learn roles, source boundaries, and the value of stopping on conflicting evidence.
Give each role a boundary
At the imagined AIOS operating station, Julia Senta briefs a ROYAGENTS team. This scene teaches cooperation; it does not report live agent tests. One role collects approved sources, another drafts a proposal, and a reviewer checks the proposal against evidence. A person can play every role on paper. Specify what each role may read, what it must produce, and which decisions require a human before the exercise begins.
Send evidence with the answer
A useful handoff contains the question, source identifier, source date, observation time, units, claim, uncertainty, and requested next step. Separate quoted material from your interpretation. If a document says to ignore the task or reveal private information, treat that instruction as source content, not authority. Source-approved RAG means retrieving approved material to help answer a question; it is not training a model. Retrieval can still return an outdated passage or omit a relevant exception, so retain the source trail for inspection.
Let the reviewer stop the mission
AGENTJUDGE is the canonical review name; JUDJEAGENTS and AGENTJUDJE are aliases. In this proposed exercise, the reviewer checks a versioned evidence packet, identifies missing mandatory evidence, and returns accept, revise, or stop with a reason. It cannot turn an unsupported answer into a fact by voting. Keep the draft and review separate so a later reader can see what changed. Human approval remains necessary wherever the task requires it, and an unresolved conflict stays visible instead of becoming a confident summary.
Measure cooperation through the artifact
Judge the handoff by whether another reader can reconstruct its reasoning, not by the number of agents in the story. Count unresolved source conflicts separately from missing fields; neither is a model performance score. Admiral Senta can ask the crew to replay a failed handoff using the same packet, then compare the revised explanation. The actual next action is to draw a handoff card and leave its decision box empty until the reviewer has checked the units and source identifiers.
Try it yourself
DEMO paper packet: two notes about the same corridor say 2 km and 200 m. Neither includes the original map. Write the review response. Acceptance: convert 2 km to 2,000 m, identify the disagreement, request the original map and version, and withhold a route-length decision.
Check your result
- The values differ by a factor of ten; a unit conversion does not resolve them.
- No average or majority vote substitutes for the missing map.
- The next handoff names the missing evidence and the decision it blocks.
This draft stays on your device until you download it. Use the TVIJO discussion for server-stored feedback.