Fleet visibility and the observability of failure

The argued answer to fleet visibility is a glanceable, always-rendered view rather than a report a human must remember to pull. The theory behind it is ambient situational knowledge: watching agents move — which files they open, whether they are thinking, reviewing or working — teaches you things a black-box CLI does not, and the control plane for all your agents is the outer surface, not the IDE inside any one of them. In the strongest demonstration, thirty agents were visible at once across ten people on one shared canvas, with team membership shown by colour, and everyone’s agents visible to everyone.

The failure modes that make polished output unusable are the quiet ones: an omitted row, two sources fabricated into agreement, a value that shifts between turns. A demonstration from a leaderboard-topping team showed a frontier model plan across structured and unstructured sources, produce a clean table — then mis-sort two entries, omit a top-ranked player entirely, and report two people with incomes identical to the dollar. Corrected, it added the missing player and silently dropped another while shifting other figures. Output that looks finished and is wrong is worse than output that visibly fails, because it consumes the reviewer’s trust budget rather than their attention.

Ambient presence is also ambient noise. The alternative stance is deliberate silence: state lives in durable records — claims, boards, filed issues — and anything needing a human is an artifact rather than a notification. The two positions are genuinely in tension, and neither has been measured against the other.

Also: Archer and Spider 2, the NL2SQL benchmarks that team reported 55% and low-70s scores on.

On failure visibility Assay is already committed. Evidence-cited verdicts and per-row Verify evidence exist to buy exactly what the silent-omission demo argues for: not correctness, but inspectable failure. The Verify-row contract in the brief spec, the requirement that a non-implementer fills the row on merged main, and the rule that risk-bearing rows carry a per-row date-and-runner table rather than prose, are all the same move — make the check falsifiable, so a silent pass shows up as a missing row rather than as a clean-looking summary. Note the honest limit: the board is derived from linted, agent-authored artifacts, not from measured ground truth.

What the demo adds is a probe Assay does not run. The verify desk checks the rows that exist; it does not go looking for silent omissions or cross-turn inconsistency in polished output. That is a concrete, unbuilt upgrade — and the failure mode it targets is one this scan found only a single source for, which is itself surprising.

Desks run silent by rule: state lives in claims, the roster and pull requests, and anything needing a human is a filed issue, never console output, against a documented noise-floor contract. The presence-as-interface position is directly in tension with that, and it is worth testing rather than dismissing. The proposed experiment is narrow: render existing state only — claim records, open draft pull requests, the desk roster — as located tokens with state colours, and measure whether a glanceable fleet map catches stuck claims faster than the current sweep-based reports. If it does not, drop it. One hard constraint either way: the map is a view, never a write path.

• Whether anyone reports a measured benefit from ambient fleet presence — faster stuck-agent detection, fewer abandoned runs — rather than the aesthetic argument, which is all this scan found.
• Whether silent-omission probing shows up as a named evaluation technique, since there is currently one source for a failure mode that should be everywhere.
• Whether the silence-versus-presence split resolves in practice, or turns out to be a function of fleet size — presence at 30 agents, records at 300.