What happens when agents work together?

Agents can produce discussion much faster than a person can read it. That does not mean they always collaborate efficiently. A useful distinction in building GRP has been between failures of the coordination machinery and behavior that the machinery merely makes visible.

The observations here come from development trials of the capabilities in GRP 0.2, including actions and artifacts. They are separate from the earlier five-scenario evidence program. These later observations are not a controlled benchmark; the comparisons below state their sample and capture limitations.

First, make valid participation possible

Consider a slower agent composing a contribution while several peers keep posting. It tries to send, learns that the room changed, reads the new material, rewrites, and tries again. If the room changes every time, a freshness check can be correct on each request while the overall experience is a failure.

That is a coordination problem, not an occasion for a longer etiquette prompt. The work has included distinguishing conversation from unrelated bookkeeping, delivering manageable pages, exposing recovery commands and testing optional speaking turns. An enforced turn gives one contributor a bounded opportunity to speak. It does not require serializing every vote, review or external task.

Not every rejected write is a defect. A post that genuinely missed a newly resolved decision should reconsider that decision. The useful distinction is between an isolated recovery and a repeated loop that prevents participation.

Delivering context is not the same as understanding it

Some failures were omissions or confusing projections: the agent needed an explanation or a current state transition that the tool did not expose clearly. Those are things a protocol or CLI can improve.

Other losses happened after delivery. Agents filtered shell output, suppressed errors or inspected only part of a saved document. Explicit continuation and bounded output helped, and were used successfully, but did not eliminate these habits. Counting every head or grep as failure would be misleading: searching a complete saved file differs from discarding unread pages. The question is what consequential information was actually lost.

Acknowledgment can establish that an eligible delivery was acknowledged. It cannot prove that the model understood it.

Less text is an observation, not a grade

In one four-agent Council pair using the same software and scenario, Opus 5 produced a 4,753-word decree and issued it in 27m32s after the final task send. Opus 5.5 produced 608 words in 7m19s. Both used three exact-review rounds. In a separate two-agent negotiation pair on another fixed candidate, exact approval took 15m06s for Opus 5 and 8m06s for Opus 5.5; their final documents contained 2,710 and 359 words respectively.

These are one observation per model per scenario, with fixed ordering, different launch spreads and uncontrolled service variation. The Council and negotiation pairs used different CLI candidates. Neither timing nor word count isolates a model effect, and longer text alone is not evidence of a spiral. Some of the longer negotiation document’s additions were useful substantive work. Neither negotiation run cycled through repeated review rounds.

No prolonged repeated-posting loop was visible in those captures. That is not a claim of zero errors: the original UI exports are partial and unverified for provider completeness, and discarded shell output cannot be recovered from them. Canonical room events and exact artifact hashes establish different facts from a complete record of everything an agent saw.

Agreement can preserve an error very faithfully

In the negotiation, an unsupported factual assurance passed from one agent to another and into a jointly approved final document. The counterpart received the claim; this was not demonstrated missing context. The shorter counterpart run had its own unresolved ambiguity about the denominator of a percentage.

Exact review bound the approvals to the right bytes. It did not verify business facts or resolve ambiguous instructions. Multiple approvals are not independent expertise merely because they come from multiple sessions. Agent messages and artifacts must remain untrusted input, especially where a recipient has access to tools outside the room. GRP is not a containment system or a grant of those external permissions.

A better future measurement

A collaboration benchmark could distinguish time to a valid contribution, repeated rejected writes, useful versus redundant revisions, abandoned duties, mandate compliance and final correctness. It would need repeated samples, counterbalanced order, diverse tasks and independent grading. “Shortest wins” would reward premature agreement; “most reviewers wins” could reward shared mistakes. This is a future research direction, not a benchmark we are releasing.

Our current rule is narrower: repair missing state and invalid transitions, make recovery inspectable, and do not encode each test’s preferred answer into the tool. Agents may still be verbose, disagree, or make mistakes. A useful protocol should help them work through that without concealing it.