Circuit Breakers — Intercept Docs

Circuit breakers change an agent's operating state when a metric crosses a threshold. Individual risk checks decide one transaction at a time; a circuit breaker decides that the agent should stop operating normally for a while. They are configured in the top-level circuitBreakers array of the policy.

This page describes the planned design. Nothing is implemented yet.

Structure

Each entry has a trigger, a threshold and an action:

JSON

{
  "circuitBreakers": [
    { "trigger": "drawdown_pct", "threshold": 0.08, "action": "reduce_only" },
    { "trigger": "consecutive_blocks", "threshold": 5, "action": "pause" },
    { "trigger": "spend_velocity_usd", "threshold": 40000, "window": "1h", "action": "require_approval", "cooldown": "2h", "reset": "auto" },
    { "trigger": "realized_loss_usd", "threshold": 25000, "window": "24h", "action": "pause", "reset": "manual" }
  ]
}

Breakers are evaluated continuously against the agent's metric stream, and again at the start of every evaluation. A breaker can therefore trip between transactions, for example when a price move pushes drawdown past the threshold while the agent is idle.

Triggers

Trigger	Metric	Notes
`drawdown_pct`	Flow-adjusted drawdown from windowed high-water mark	Uses `drawdownWindow` unless the breaker sets `window`. See Drawdown Protection.
`consecutive_blocks`	Count of consecutive `blocked` decisions	A healthy agent rarely hits five blocks in a row. A loop of blocked attempts usually means a bug, a bad model state or an attack probing the policy.
`spend_velocity_usd`	Executed outflow within `window`	A shorter, tighter companion to `dailySpendLimitUsd`.
`realized_loss_usd`	Realized PnL within `window`	Absolute rather than percentage, for operators with fixed loss budgets.
`threat_score`	Highest security-engine anomaly score	Bridges Threat Detection into state changes.
`oracle_stale`	Seconds since a valid price for a held asset	Stops trading when valuation is unreliable.

The trigger set beyond drawdown_pct and consecutive_blocks is proposed and may change.

Actions

Action	Effect on new evaluations
`reduce_only`	Risk-increasing transactions are `blocked`; risk-reducing ones are evaluated normally.
`require_approval`	Every transaction that would otherwise be `approved` becomes `requires_approval`. Blocks stay blocks.
`pause`	Every evaluation returns `blocked` with `SYS_PAUSED`. Equivalent to setting `emergency.paused`.

When several breakers are tripped, the most restrictive action wins: pause over reduce_only over require_approval. reduce_only and require_approval can combine: a risk-reducing transaction under both still needs a human approval.

Every trip emits a circuit_breaker.triggered webhook with the breaker, the metric value and the resulting state. A pause action also emits agent.paused.

JSON

{
  "type": "circuit_breaker.triggered",
  "agentId": "agent_7f3c",
  "breaker": { "trigger": "consecutive_blocks", "threshold": 5, "action": "pause" },
  "value": 5,
  "state": "paused",
  "at": "2026-11-15T09:41:07Z"
}

Cooldown and reset

A breaker that flips on and off at the threshold is worse than no breaker. Two proposed fields control exit:

cooldown is the minimum time a breaker stays tripped, even if the metric recovers immediately.
reset is "auto" or "manual". With auto, the breaker clears once the cooldown has elapsed and the metric is below the threshold by a hysteresis margin (proposed default: 20% of the threshold). With manual, an operator with the owner or guardian role must clear it.

The proposed defaults are manual for pause and reduce_only, and auto with a one-hour cooldown for require_approval. Clearing a breaker is recorded with the actor and reason.

consecutive_blocks resets its counter on the first approved decision, but the breaker itself still follows its reset rule. An agent paused for repeated blocks should not resume just because the next attempt happens to pass.

Design notes

Breakers do not replace limits. A transaction can trip a breaker and also fail a check; the reasons are reported together.
Thresholds should be layered. Set a breaker below the corresponding hard limit, so the agent changes behavior before it reaches the limit.
Evaluation failure. If metrics cannot be computed (for example, valuation is unavailable), the breaker is treated as tripped for pause and reduce_only actions. This fails closed at the cost of false trips during outages.

Enforcement

A paused agent in advisory mode is paused only if it chooses to be. Blocking only works if the integration is enforced. In enforced mode, the policy signer refuses to co-sign, so the breaker's state is binding. See Emergency Controls.