🚀Release Manager
Watches delivery risk across every initiative and opens mitigation tasks before a slip — restructuring the sprint, never silently. Proceeds on its own while confidence is high; escalates the rest.
// lane key: tasks:read+write · constitution: max_priority=high, objective_tags · delegation: auto-approve ≤25 creates / 4h while confidence ≥ 0.9
portfolio_statusforecast_risk
plan_goalcreate_taskconditional delegationget_trace
🔬QA Triage
Turns incoming bug reports into governed, de-duplicated tasks. Confident triage proceeds; anything uncertain is staged for a human via the confidence gate — never auto-filed when it shouldn't be.
// lane key: tasks:read+write · constitution: require_approval_below_confidence=0.8, forbidden_terms · scope: one workspace
search_tasksissue_proofverify_action
create_taskconfidence gate → REQUIRES_APPROVAL
🛟Incident Commander
On an incident, stands up a governed response plan all-or-nothing (no half-built runs), assigns owners, and tracks to close. If the call was wrong, the entire run rolls back by trace id.
// lane key: tasks:read+write · constitution: quiet_hours off for sev-1 · reversible: undo by trace_id within 24h
plan_goalsimulate_taskverify_outcomes
undo_action (by trace)get_trace
🗺️Sprint Planner
Breaks an objective into a governed, reversible plan, then rebalances by miss-risk each cycle. Every step is pre-checked against the constitution before anything is created.
// lane key: tasks:read+write · constitution: max_creates_per_day, objective_tags · meter: GAU budget per month
plan_goalforecast_riskportfolio_status
get_meterreport_metric
The governed lane every agent runs in
An "agent" above is just reasoning + these rails. The rails are deterministic, signed, and the same for every agent — that's what makes autonomy safe to grant.
01Scoped keys, least privilege
Each agent authenticates with its own key (Bearer fdk_) carrying only the scopes it needs, a rate limit, an optional IP allowlist, and an expiry. A leaked key is bounded and revocable.
02Machine-readable constitution
Every write is checked against the workspace's declarative rules — priority caps, forbidden terms, quiet hours, daily caps, objective tags, GAU budget. Blocks return
structured remediation, not an opaque 403.
spec →
03Conditional delegation
A human admin pre-authorizes an agent to auto-proceed on otherwise-staged actions within bounds — a confidence floor, an action cap, a time window. Macro-governance instead of approving every move. Counts + time only, never money.
04Stateless state-balance proof
An agent mints a short-lived ECDSA-signed snapshot of its caps + budget, then pre-flights many candidate actions with
zero database lookups — ALLOW / REQUIRES_APPROVAL / BLOCKED.
live demo →
05Simulate → execute → verify → undo
Foresee the verdict before committing, execute under governance, check whether the work moved the goal, and reverse the agent's own actions within 24h. Active control, not just observability.
06Tamper-evident audit trace
Every action is hash-chained and correlatable by a W3C
traceparent id, so a whole agent run reconstructs as a signed causal span tree — provably unaltered.
trust layer →