Every payload a host hands back from a delegated step names the model that produced it and the number of agents that ran, and the ingesting verb refuses one that does not

Press Release

Seeded from a quoted-text intent capture. Expand into the full press-release narrative before planning.

Why This Matters

Every payload a host hands back from a delegated step names the model that produced it and the number of agents that ran, and the ingesting verb refuses one that does not

Mechanism

Prompted (the claim-recording gradient): why the authors expect this to work, as a falsifiable "we expect X because Y" — not the outcome restated. Replace this line with the claim, or with the exact token None stated. alone on its line to record the claim as considered and declined.

Scope Conditions

Required (the claim-recording gradient): the population, platform, scale, or assumptions this claim holds under, one per top-level bullet — abcd intent plan stamps each with a persistent identity. Replace this line with those bullets, or with the exact token None stated. alone on its line.

Acceptance Criteria

Required (the itd-1 discipline): add at least one Given-When-Then bullet describing the verifiable bar for "shipped" before this draft can be planned.

SOTA

Surveyed on 2026-10-23; sources, evidence tiers and the local fit risks are in the autonomous-loops SOTA note.

• The field names exist, on different spans. OpenTelemetry's GenAI conventions (status: Development) define gen_ai.request.model (on the create-agent span, "the exact name of the model requested"; on the invoke-agent span Recommended only where one model applies), gen_ai.response.model ("the name of the model that generated the response", defined on the model-call spans and not listed on the invoke-agent span at all), gen_ai.provider.name, gen_ai.agent.id, gen_ai.agent.name, gen_ai.conversation.id and gen_ai.usage.input_tokens / output_tokens, with input tokens including cached ones (OTel GenAI spans and agent spans). A per-agent gen_ai.response.model, and the agents list itself, are abcd's own composition of OTel names, not an OTel structure.
• The host's run already carries the facts. system/init reports the session's model; the result's modelUsage is keyed by model name and counts sub-agent requests where usage does not; every sub-agent message carries a parent_tool_use_id from which the nesting tree is rebuilt; SubagentStop carries agent_id and agent_type but no model; cost figures are client-side estimates (Claude Code and Agent SDK docs). So the per-model totals come from the final result, but the per-agent served model and tokens can only be rebuilt from the stream (stream-json messages with parent_tool_use_id, each carrying its model and usage), not from the result alone; and after a process crash the final error_during_execution result "may carry zeroed usage, total_cost_usd, and modelUsage", so a run that did work can hand back an empty agent list, and the refusal must say what it does with a zeroed result rather than treat it as no work.
• Provenance words. W3C PROV-O: an entity wasAttributedTo an agent, an activity wasAssociatedWith an agent, an agent actedOnBehalfOf another, and SoftwareAgent is a kind of Agent; the human the run acts on behalf of is the attribution rule's author.
• Served is not asked. The 2026-10-16 note measured the spread between provider pairings; the build verb's scope condition already says the model is recorded as the runner reported it and cannot be verified by the binary.

Path 1 for the names, Path 2 for the mechanism: the payload carries the OTel attribute names verbatim (gen_ai.provider.name, gen_ai.request.model, gen_ai.response.model) and an agents list keyed by gen_ai.agent.id with each agent's served model and tokens, the per-agent gen_ai.response.model being abcd's extension of a convention whose agent span does not define it; the agent count is the number of distinct agent ids that ran, the top-level one included; the ingesting verb refuses a payload with no gen_ai.response.model or an empty agent list, and a zeroed crash result is recorded as "unknown" with the crash named rather than refused as no work; the per-agent fields are read from the stream, the per-model totals from the result; the payload pins the convention version because the status is Development. Names only, no OTel SDK, no Go dependency. The independent fit-challenge runs at plan time.

Open Questions

• (SOTA 2026-10-23) Whether "the number of agents that ran" counts distinct agent ids (the top-level one included) or only sub-agents, and whether a forked skill with a forked-command- parent counts as an agent.
• (SOTA 2026-10-23) Whether the payload records both the requested and the served model, and refuses only when the served one is missing.
• (SOTA 2026-10-23) Which convention version the payload pins, and what the ingesting verb does with a payload pinned to a newer one.

Audit Notes

Empty. Populated by intent-auditor when intent moves to shipped/.