When should an always-on agent speak up unprompted, and does anyone's answer transfer to Auto Mode?
The question
"What decision rules govern when an always-on AI agent should speak up unprompted in a shared/async channel, and do any of them transfer to RDCO's own Auto Mode escalation design?"
Every's platform lead named this as the one problem his team has not solved for @Every in Slack ([[2026-10-20-every-shared-agent-permissions-architecture]]), while RDCO has been running an undocumented answer on iMessage for seven months.
What we already know (from the vault)
- RDCO's trigger is a four-category content taxonomy, and it is founder-ratified. Channel messages are reserved for "decisions awaiting his judgment, surprising findings, actual blockers, or things he asked me to surface" and explicitly not for "status reports of expected work, asking confirmation on choices I can defensibly make myself, narrating what I'm about to do" ([[feedback_auto_mode_signal_to_noise]]). The founder's own words on 2026-04-20: flipping to auto mode "greatly reduced the number of times you have to come back to me for approval... a noticeable improvement in the signal to noise ratio." A 2026-04-24 correction sharpened it: of four items Ray surfaced as DECISION NEEDED, three were reversible actions it should have just executed ([[feedback_distinguish_decision_from_action]]).
- RDCO separates the speech decision from the action decision, and only the action half is mechanically enforced. Deploy and production-write denials are a hard gate after one denial, and a channel message explicitly does not count as authorization because a chat can be spoofed ([[feedback_automode_classifier_hard_gate]]). Nothing equivalent gates speech: whether to send a bubble is a model judgment every time.
- RDCO prices its own asks, which is unusual. When Ray sends an ask, a 60-minute cooldown starts; past it, the
/open-threads-checkcron parks the thread with a 🅿️ marker and the morning brief resurfaces it. Off-hours (9pm-9am ET) do not auto-park, deliberately ([[feedback_open_threads_channel_state]]). The discipline note says "don't over-park... false-positive parking creates channel noise." - The mechanical-over-instructed case is already evidenced internally. Classifier block rate on dangerous commands stays flat with session length while unaided human vigilance decays from roughly 17% to 5% after fifty-plus prompts ([[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]]), which is the empirical argument behind [[2026-10-25-coded-vs-prompted-permission-gates]].
- One standing rule constrains the obvious fix. Ray is forbidden from managing the founder's time, attention, or focus ([[feedback_no_babysitting]]). Any urgency-gating scheme that infers whether he is interruptible right now runs straight into it.
What the web says
- Only one source states a general decision rule for unprompted agent speech. Zhu, Wang, Xiao and Shen (Carnegie Mellon), "Design Principles for Human-Agent Interaction," arXiv:2606.20630v1, 28 May 2026, Principle 7 "Manage Proactive Initiative": "Time proactive interventions ... to match task urgency and context, and default to responding rather than proactive intervention." The principle adds: "Minimize non-urgent interruptions from agents because users tend to have low tolerance for them."
- That rule's backing is expert judgment, not measurement. The paper is a self-labeled position paper; its evaluation is a heuristic walkthrough by 3 evaluators across 9 deployed agents, ~1.5h each, rated -2 to +2. No inter-rater reliability, no significance testing, and no per-principle mean is reported for Principle 7 specifically. Its own citations are to voice-assistant and guidance studies (Edwards et al. on urgency moderating interruption acceptance; Cai et al. finding high proactive guidance raises self-awareness but lowers perceived usefulness).
- Anthropic's position reaches us only secondhand, and it is a slogan plus an undefined number. VentureBeat (Michael Nuñez, 24 Aug 2026) reports Anthropic's Claude Slack agent update: Anthropic wrote that "an annoying agent is worse than an unhelpful one," claims the update is "roughly 30% better at deciding when - and, critically, when not - to jump into a conversation unprompted," and reportedly removed the lightweight per-message classifier that made a binary respond/don't-respond call. No baseline, metric definition, or sample size is given for the 30%. The claim that Claude goes dormant in channels where it has nothing to add is the reporter's summary, not a quoted vendor rule.
- Anthropic's enumerated outcome set is the one genuinely new structural idea. Per the same report, Claude now chooses among four outcomes: reply inline, open deeper work in a thread, route the message into an existing workstream, or say nothing. No criteria for choosing among them are described. Trigger logic is asserted, not specified.
- The only shipped mechanism with real detail delegates the rule to the customer. Adapt's changelog (13 May 2026, unsigned vendor post) ships per-channel and per-thread trigger modes - @-mention-only, Smart, Always respond, with thread overriding channel - and "Nothing changes until you explicitly opt a scope into one of the new modes." Smart mode runs a fast judge model that will "decide in milliseconds whether each new message warrants a reply" against a plain-English policy the user writes. The vendor states no substantive speech rule of its own, and publishes zero evidence: no accuracy figures, no latency measurement, no study.
- Every admits it has no rule. [[2026-10-20-every-shared-agent-permissions-architecture]] documents the fixes for conversations, connections and permissions, and names "when should the agent speak up unprompted" as the open question. Its one published heuristic, Will Gathright's "how hard would this be to undo?", governs act-versus-ask, not speak-versus-stay-silent.
- The published literature on act-versus-ask is far ahead of the literature on speak-versus-silent. Across the vault's always-on-agent coverage, every vendor treats approval-gating of irreversible action as non-negotiable ([[2026-10-25-always-on-personal-agent-product-category]]), and Every's own team wrote Dot an explicit boundary - "Don't send anything without my approval" ([[2026-10-23-every-openai-dots-guide-always-on-agents]]). Nobody has published the equivalent for speech.
Convergences and contradictions
- Four independent sources converge on default-to-silence and nothing else. CMU's "default to responding rather than proactive intervention," Adapt's @-mention-only default, Anthropic's "an annoying agent is worse than an unhelpful one," and RDCO's reserve-for-four-categories all land in the same place. That convergence is real and worth banking. What does not exist is any shared account of what lifts the agent off that default: CMU says task urgency, Adapt says whatever the customer wrote down, Anthropic says the model's own judgment over full channel context, RDCO says a fixed category list. Those are three incompatible loci for the decision - in the human's written policy, in the model, or in a hard-coded taxonomy - and no source compares them.
- Anthropic moved in the opposite direction from RDCO's own evidence. Deleting an explicit classifier in favor of in-model judgment is precisely the instructed-over-mechanical trade that [[2026-10-25-coded-vs-prompted-permission-gates]] and the approval-fatigue numbers argue against. The only evidence offered for the move is one undefined 30%, reported by a third party. Treat it as a vendor direction, not a validated finding.
- RDCO holds a mechanism nobody else has published, and lacks one everybody else assumes. The 60-minute cooldown plus 🅿️ parking ([[feedback_open_threads_channel_state]]) prices the agent's own asks and routes unanswered ones into a daily digest; no source found here does anything comparable. Conversely, every published rule gates on recipient state or urgency, and RDCO gates purely on item category, because [[feedback_no_babysitting]] forbids modeling the founder's attention.
Synthesis for RDCO
The honest headline is that the field has not settled this, and RDCO should stop treating its own answer as the ad hoc one. Exactly one source in this sweep states a general rule for unprompted agent speech, it is a CMU position paper whose backing is three evaluators' heuristic judgment, and its own evidence base is borrowed from voice-assistant studies where an interruption is synchronous and unignorable. Async text is a different cost structure. Adapt ships a real configurable mechanism with no evidence and no opinion. Anthropic ships an opinion and an undefined number. Every ships an admission. Against that, RDCO has a trigger taxonomy the founder ratified in his own words, a correction log that sharpened it after a specific miss, and a cooldown mechanism that is genuinely novel. The comparison runs in RDCO's favor, and the correct update is to raise confidence in what is already running rather than to go shopping.
Two things should not change. First, the four-category trigger stays. Replacing a founder-validated taxonomy with Adapt's write-your-own-policy model would trade a rule tested against real corrections for a configuration surface with no published evidence behind it, and it would hand the founder a prompt-writing chore in exchange for flexibility a single-recipient DM does not need. Second, do not follow Anthropic in deleting the explicit gate in favor of model judgment. RDCO's own data says the mechanical check holds its rate while judgment decays, and the classifier hard-gate is load-bearing specifically because a channel message cannot authorize a production write. Softening the speech side toward in-model discretion while the action side stays mechanically gated would be a coherent split; softening both would not.
One thing is worth adopting, cheaply and reversibly. RDCO's speech decision is currently binary: send a bubble or stay silent. Anthropic's enumerated four outcomes - reply inline, open a thread, route into an existing workstream, say nothing - name a middle that RDCO half-practices and never decided. Today "brief iMessage, link to HQ" and the morning-brief digest both function as that middle, but they are chosen by habit rather than at speech time. Making it an explicit four-way choice (bubble now / append to an existing open thread / file to vault or HQ and surface in the brief / nothing) costs one paragraph in the Auto Mode rules and would cut fresh bubbles for items that are real but not urgent. Frame that as one vendor's enumeration reported secondhand, not as industry practice, and adopt the shape without the claim.
The urgency question should be named and then left alone. CMU's rule gates on task urgency and the recipient's current context, which is exactly the variable [[feedback_no_babysitting]] forbids Ray from inferring. The resolution already exists and looks accidental: the 9am-9pm parking window and the off-hours no-park carve-out are a coarse, mechanical proxy for recipient availability that requires no attention modeling at all. That is the right trade for a single-human channel, and it should be written down as a deliberate substitution so nobody later "fixes" it by building attention inference. Finally, the shared-channel half of the question mostly does not transfer. The ambient-spam problem, bystander externalities, per-channel scoping and re-tagging etiquette are all artifacts of a multi-party room; RDCO runs a single-recipient DM. What transfers is the attention budget of one person, and that is the narrower, more tractable problem RDCO is actually solving.
Why this is in the vault
This is the first vault doc to compare published decision rules for unprompted agent speech, and it closes the gap flagged in [[2026-10-20-every-shared-agent-permissions-architecture]] by establishing that RDCO's Auto Mode trigger is better evidenced than any shipped alternative found - which converts a suspected weakness in the channels agent ([[project_channels_agent_setup]]) into a defensible position, and gives the agents-in-production credibility thesis an original mechanism (the ask cooldown) to write from rather than a restatement of vendor posts.
Open follow-ups
- Does Anthropic's own primary source (product doc, engineering post, or Slack app documentation) describe the trigger logic that replaced the deleted per-message classifier, and define what the 30% figure measures? Only secondhand press was reachable in this sweep.
- Has any controlled study measured interruption cost for async text channel agents, as opposed to the voice-assistant literature (Edwards et al.) that CMU's Principle 7 rests on? The cost structures differ enough that the transfer should not be assumed.
- Is there published evidence for any specific cooldown or digest interval on an agent's unanswered ask, or is RDCO's 60 minutes arbitrary? The number has never been tested against anything.
- Do any shipped agents gate unprompted speech on recipient state (calendar, do-not-disturb, typing presence) rather than message category, and has either approach been measured against the other?
- For a single-recipient async channel, does batching into one digest beat per-item interjection on response rate or decision latency? Both CMU and Adapt assume multi-party rooms, so their restraint findings may not carry.
Related
- [[2026-10-20-every-shared-agent-permissions-architecture]]
- [[2026-10-25-always-on-personal-agent-product-category]]
- [[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]]
- [[2026-10-25-coded-vs-prompted-permission-gates]]
- [[2026-10-23-every-openai-dots-guide-always-on-agents]]
- [[2026-10-29-every-first-week-ai-coworker]]
- [[2026-10-30-every-marx-alienation-coding-agents]]
- [[2026-10-23-rdco-auto-mode-classifier-miss-profile]]
- [[project_channels_agent_setup]]
- [[feedback_auto_mode_signal_to_noise]]
- [[feedback_distinguish_decision_from_action]]
- [[feedback_automode_classifier_hard_gate]]
- [[feedback_open_threads_channel_state]]
- [[feedback_escalation_note_standard]]
- [[feedback_no_babysitting]]
Sources
Vault
06-reference/2026-10-20-every-shared-agent-permissions-architecture.md- Willie Williams (Every), names the unresolved question06-reference/concepts/2026-10-25-always-on-personal-agent-product-category.md- observe-vs-act dial, approval-gate convergence06-reference/research/2026-08-13-auto-mode-approval-fatigue-channels-session-pattern.md- flat classifier rate vs decaying human vigilance06-reference/concepts/2026-10-25-coded-vs-prompted-permission-gates.md- mechanical over instructed enforcement06-reference/2026-10-23-every-openai-dots-guide-always-on-agents.md- Every's team-written boundary for Dot06-reference/2026-10-29-every-first-week-ai-coworker.md- Grant Harvey's team-touching / private-drafting / human-judgment split- Memory:
feedback_auto_mode_signal_to_noise,feedback_distinguish_decision_from_action,feedback_automode_classifier_hard_gate,feedback_open_threads_channel_state,feedback_escalation_note_standard,feedback_no_babysitting
Web
- Haiyi Zhu, Canwen Wang, Qing Xiao, Hong Shen (Carnegie Mellon), "Design Principles for Human-Agent Interaction," arXiv:2606.20630v1 [cs.HC], 28 May 2026 - https://arxiv.org/pdf/2606.20630 (preprint / position paper; heuristic evaluation, n=3 evaluators, 9 agents)
- Michael Nuñez, "Anthropic's new Claude Tag update lets its Slack agent read the full conversation - and jump in unprompted," VentureBeat, 24 Aug 2026 - https://venturebeat.com/orchestration/anthropics-new-claude-tag-update-lets-its-slack-agent-read-the-full-conversation-and-jump-in-unprompted (secondhand press; Anthropic's own post not reachable in this sweep)
- Adapt, "Proactive agent mode in Slack," changelog, 13 May 2026 - https://adapt.com/changelog/proactive-agent-slack (unsigned vendor changelog; no evidence published)