Bounded Planner / Executor / Critic
The Decision plane runs as a bounded triad: the Planner produces a typed plan; the Executor runs only verified steps via the Tool Gateway; the Critic verifies and scores. Subagent lanes isolate parallel work; durable background sessions resume by checkpoint. Replay is a property of the system.
Planner, Executor, and Critic operate as separate lanes with checkpoints, verified retries, subagent isolation, and a durable replay path.
- 01Planner
Produces typed steps, dependencies, and declared risk.
- 02Critic verify
Rejects unsafe or underspecified plans before action.
- 03Executor
Runs verified steps only through approved tool envelopes.
- 04Critic score
Accepts, retries, replans, or escalates from evidence.
- 05Checkpoint
Persists session state for replay and resume.
Three roles, three artifacts
Plan, transcripts, verdict — each independently auditable. The Critic re-derives its verdict offline during replay without re-executing tools.
Anti-pattern prevented
Single agentic loop: conflates planning, execution, and judgment, making failures uninterpretable. The bounded triad fixes that.
Metrics that matter
Plan-verification pass rate, retry/recovery rate, escalation rate by risk tier, mean time to safe completion, loop-guard trip rate.
Typed plan with decision checkpoints
Plans are typed JSON with steps, dependencies, declared approval_mode per step, and decision checkpoints that resolve to a Decision Spec from the Decision Catalog.
{
"plan_id": "plan_refund_01",
"intent": "support.refund",
"steps": [
{ "id": "s1", "tool": "adp_orders.lookup", "params": { "order_id": "ord_881" } },
{ "id": "s2", "tool": "adp_policy.eval", "depends_on": ["s1"] },
{
"id": "s3",
"tool": "adp_payments.issue_refund",
"depends_on": ["s2"],
"approval_mode": "destructive",
"requires": ["GATE_FINANCE_APPROVAL"]
}
],
"decision_checkpoints": [
{ "decision_id": "support.refund.execute", "after_step": "s2" }
]
}