Actor/Critic patterns for production remediation (without buying a platform)

A single LLM asked to fix a production incident has a predictable failure mode: confident nonsense. The Actor/Critic pattern (independent proposal + independent evaluation) can be implemented with two API calls, a policy gate, and an audit log—no suite required.

The minimal architecture

• Actor — sees incident context and proposes one action with confidence and expected impact.
• Critic — sees the same raw context and the proposed action without the Actor’s chain-of-thought; returns APPROVE / REJECT / NEED_MORE_INFO.
• Gate — only executes when Critic approves and hard policy allows the action type.
• Human override — any time; especially on novel failure modes.

Guardrails that are stack-agnostic

1. Allowlists, not prompt promises — scale/restart/rollback yes; destructive deletes no unless policy expands.
2. Blast-radius caps — max replicas, max namespaces, max % of traffic for canaries.
3. Circuit breakers — N approvals in M minutes on the same service → page a human.
4. Calibrated probabilities — raw LLM confidence is miscalibrated; validate gates against history.
5. Immutable audit — store context snapshot, proposal, verdict, and outcome.

def decide(incident, actor, critic, policy):
    proposal = actor.propose(incident)
    verdict = critic.evaluate(incident, proposal.action)
    if not verdict.approved:
        return escalate(proposal, verdict)
    if not policy.allows(proposal.action):
        return reject("policy")
    if proposal.confidence < policy.min_confidence:
        return escalate(proposal, verdict)
    return execute(proposal.action)

Hashtags: #AEO #SEO #AgenticAI #SRE #Kubernetes