Ambient agents

Most agents are reactive: you ask, they answer. Thoth has a parallel mode — ambient agents — that wake on their own when something interesting happens.

This is what makes the agent feel like it cares about your work, rather than just responds to it.

What ambient agents do

A few examples of triggers that fire ambient agents in production:

• Morning digest — at 09:00 your local time, the agent posts a Slack DM summarizing what’s queued, what’s stale, what changed overnight
• Stale skill nudge — a skill draft has been pending approval for

  3 days, the agent gently DMs to ask if you want to ✅ or ❌

• Self-spawn follow-up — reflection set a next_check_at timestamp; the agent posts back into the original thread when the time arrives
• Status incident announcement — status.claude.com reports an incident; the agent posts to a designated channel

The agent isn’t always doing this — it’s bounded by a daily budget envelope (default $1/day, configurable). When the budget exhausts, ambient activity pauses until the next UTC day.

The budget envelope

The fundamental constraint is AMBIENT_DAILY_CAP_USD (default 1.0). Every ambient action — every LLM call made by a trigger — debits from a per-day pool.

Day starts (00:00 UTC):
  budget remaining: $1.00

Trigger A fires (morning digest, $0.07 LLM call):
  budget remaining: $0.93

Trigger B fires (skill-draft nudge, $0.01 LLM call):
  budget remaining: $0.92

Trigger C fires (self-spawn follow-up, $0.15 LLM call):
  budget remaining: $0.77

[5 more triggers throughout day]

Trigger Z fires ($0.20 LLM call):
  insufficient budget → trigger DEFERRED to next day

This is visible autonomy: the dashboard’s “ambient” tile shows real-time burndown. Trust earned by being legible, not by being cautious.

Trigger types

Out of the box, Thoth ships with these trigger types:

1. Cron triggers (scheduled)

Fire on a cron schedule. Used for morning digest, weekly summaries, nightly housekeeping.

- name: morning_digest
  type: cron
  schedule: '0 9 * * *'        # 09:00 daily
  budgetUsd: 0.50
  prompt: |
    Generate a morning digest for {founder}.
    Include: pending skill drafts, recent Sentry issues,
    threads idle >12h, party results since last digest.

2. Threshold triggers (debounce)

Fire when a counter crosses a threshold. Used for “skill draft has been pending >3 days.”

- name: stale_skill_nudge
  type: threshold
  watch: skill_drafts.pending_days >= 3
  cooldownHours: 168  # don't re-nudge same skill within a week
  budgetUsd: 0.05

3. Schedule-from-reflection (self-spawn)

Set during session-end reflection. The reflection writer can emit a next_check_at ISO timestamp; the scheduling poller fires at that time.

{
  "outcome": "partial",
  "next_check_at": "2026-05-12T10:00:00Z",
  "user_model_updates": {
    "U_EXAMPLE_USER": ["wants to revisit the Sentry refactor next Monday"]
  }
}

The poller fires at the timestamp by synthesizing a Slack message into the original thread:

🤖 self-spawn follow-up — reflection scheduled this check

Following up on Sentry refactor — any progress, or want me to draft the spec?

Caps:

• 5 self-spawns per thread (prevents loops)
• 60-second cooldown between fires
• /loop-stop cancels all pending

4. Status-change triggers

Fire on a state transition you watch. Built-in example: anthropic-status-monitor polls status.claude.com every 60s and fires when an incident transitions (new / updated / resolved).

- name: anthropic_status_announce
  type: status_change
  source: anthropic-status
  channel: '#ops-claude-status'
  budgetUsd: 0  # no LLM call needed; pure forwarding

Cost economics

A typical Pro-tier user with all default ambient triggers enabled spends about $0.30–$0.50/day on ambient activity:

Trigger	Frequency	Cost
Morning digest	1×/day	~$0.10
Stale skill nudge	1–2×/week	~$0.01
Self-spawn follow-ups	0–3×/day	~$0.10–$0.30
Status announce	0–1×/day	~$0 (no LLM)

The default $1/day cap is generous. Tune down via AMBIENT_DAILY_CAP_USD=0.50 if you want.

Disabling

Three levels of disable:

# Disable all ambient agents
AMBIENT_DISABLED=true

# Disable specific trigger types
AUTO_HARVEST_DISABLED=true              # stops the auto-skill harvester
ANTHROPIC_STATUS_DISABLED=true          # stops the status monitor

# Disable just the announcement (keep monitoring)
ANTHROPIC_STATUS_ANNOUNCE_CHANNEL=     # unset → no Slack posts

Anti-pattern: ambient as autonomous-action

Ambient triggers are designed for proactive surfacing, not autonomous action. Don’t write triggers that:

• Auto-merge code
• Auto-deploy to production
• Auto-respond to customer inquiries
• Auto-spend on external services

The mental model: ambient agents bring things to your attention, they don’t act on your behalf. Action requires explicit founder approval (✅ on a Slack card) per Thoth’s hard rules.

If you find yourself writing a trigger that sounds like “automatically X without asking,” you’re misusing ambient. Either:

1. Restructure as “DM the founder asking for ✅” before X happens
2. Move the logic into a skill that the founder explicitly invokes

This isn’t capability-gating — it’s a deliberate design choice. The agent that surprises you with autonomous actions earns less trust over time than the agent that surfaces options for you to choose.

Self-spawn pattern in depth

The most powerful ambient pattern is self-spawn from reflection.

Sequence:

1. You ask Thoth something complex: “How should I think about the Sentry refactor?”
2. Thoth replies thoughtfully but acknowledges uncertainty
3. Session ends; reflection runs over the transcript
4. Reflection notes: “outcome: partial; user wanted concrete answer but I gave conditional one”
5. Reflection emits: next_check_at: <next Monday 10:00 user TZ>
6. Scheduling poller stores the entry
7. Monday 10:00 comes. The poller dispatches a synthetic message into the original thread.
8. Thoth (now resumed via claude --resume <session>) sees the self-spawn prompt in the same context, with all the prior conversation, and posts a follow-up:

Following up on the Sentry refactor — I’ve been thinking. Want me to draft a spec or do you want to sketch the migration plan first?

This creates the impression of an agent that carries unfinished business across days. Which it does.

Caps + safety

• Max 5 self-spawn fires per thread
• 60-second cooldown between fires within a thread
• /loop-stop cancels all pending self-spawns
• Daily budget envelope caps total ambient spend
• All ambient actions logged to dashboard (transparent)

What’s next

• SPEC-decision-records — the v0.6 spec for full ambient- trigger audit trails
• Reflection at session-end — what feeds next_check_at
• Reactions — /loop-stop and reaction controls