An agent loop that stops says who it is asking and what it needs, and never simply halts

Press Release

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

Why This Matters

Without a rule about when agents may proceed alone, every gate is a candidate for a human wait, and the framework's premise reverses: the human becomes the bottleneck it exists to remove. The agents run autonomously and stop only to obtain a verdict.

A stop is an event with an addressee and a question, never an unexplained halt. It names the role it is asking, states what it needs, and carries the options that would answer it. Work that is a step in the loop rather than a judgement is performed, not queued for somebody.

Escalation reaches the product thinker only for what they can answer: what should exist, whether what was delivered matches it, and trade offs that change the design. Everything else stops at the facilitator, including every question about evidence, technique, and whether a criterion was verifiable at all. A stop addressed to the product thinker that they cannot answer is a defect in the stop, not a failure of the reader (adr-2609151528057260).

Every stop carries its addressee as data, so a surface can show whose question it is before anyone reads the words. The addressee is not a prefix somebody remembers to type: it travels with the stop, survives serialisation, and is what a renderer keys on.

A stop that asks a conjectural question carries its options under widen-options-never-recommend: at most three, the null answer always among them, none marked as preferred, and the non-recommendation stated rather than left to be inferred. A stop that reports a settled fact carries no such marker, because hedging a fact is false balance.

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.

• A standard envelope exists for the question half. MCP elicitation (specification 2025-11-25, which supersedes the 2025-06-18 shape) sends a message and, in form mode, a requestedSchema limited to flat primitives, with titled choices as oneOf: [{const, title}] (multi-select as anyOf inside an array; enumNames is gone), and takes a three-action answer: accept (chose and submitted), decline (explicitly refused), cancel (dismissed without choosing); servers "MUST NOT use form mode elicitation to request sensitive information", a URL mode sends the person to a page out of band with consent and a completion notification, and clients "MUST provide clear indication of which server is requesting information". It has no addressee, no judgement-versus-fact marker and no non-recommendation statement.
• The host routes that envelope today. An Elicitation hook receives mcp_server_name, message, mode, url, elicitation_id and requested_schema and may answer with the same three actions, so that "The server gets your answer and no dialog appears"; the host's AskUserQuestion offers options plus an Other row, its questions "stay open until you answer them", and an optional askUserQuestionTimeout closes one so that "Claude proceeds on its own judgment", a timer that "never starts" in a background session or while connected to remote control (Claude Code hooks and tools reference, re-read in this pass).
• The waiting state is data elsewhere too. A2A puts input-required and auth-required on the task as interrupted, non-terminal states, resumed on the same task id (A2A 1.0). LangGraph's interrupt payload is any JSON-serialisable value and the graph waits indefinitely; the OpenAI Agents SDK stores pending approvals as a durable run state, warns that deserialising one does not authenticate it, and advises a version marker (framework docs).
• Addressing a role is incident practice. An escalation policy's rules name users or schedules, not a person by hand, and acknowledgement does not resolve (PagerDuty docs). W3C PROV's actedOnBehalfOf names a facilitator relaying for a product thinker.

Path 2: a stop record of abcd's own whose question half is elicitation's envelope (message, flat schema, a oneOf of at most three const/title options with the null answer as a named option, accept|decline|cancel, and URL mode as the standard hand-off to a page) and whose header carries the addressee role, the kind (judgement or fact), the non-recommendation flag, the record concerned and a schema version; AskUserQuestion is the first renderer and the hook's three actions are the answer's values, so a web page or an MCP client renders the same record later. "None of these" is an enum value, never cancel; the answer is appended to the record, never loaded from an unauthenticated snapshot. No Go dependency. The independent fit-challenge runs at plan time.

Open Questions

Decided 2026-08-29. The addressee is first-class data on the stop, not a rendering convention, because the terminal is not the surface that matters most and a convention does not survive the trip to a surface abcd owns.

Decided 2026-08-29 (product thinker). A stop nobody answers waits indefinitely by default and stays visible as waiting on the product thinker. Escalating to the facilitator after a period is available but is not the default, and it is configured where the rest of the facilitator's involvement is configured (itd-174), so a repository that wants the work to keep moving can say so and a repository that does not is never overridden by a timer it did not set.

• (SOTA 2026-10-23) Whether decline (an explicit refusal to answer) and cancel (a dismissal that leaves the stop open) are both kept as answer values, or only one, given that the null option already covers "none of these".
• (SOTA 2026-10-23) Whether the host's answer-for-the-user path (the Elicitation hook, the question timeout) is ever allowed on a stop of kind judgement, or only on a stop that reports a fact.

Audit Notes

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