Autonomous ops agent
Plan, verify,
switch.
operator runs a plan-execute-verify loop that never trusts an executor's own claimed_success — a separate function recomputes pass/fail from acceptance criteria alone. On failure it switches strategy, never repeats the plan that just failed.
operator · plan-execute-verify ops loop · [higgsfield: noir gold branching decision tree, one dead path, one lit recovery path] · agent engineering
The failure mode
The success that wasn't
“The agent said it succeeded. It didn't. We only found out during the demo.”Product lead · internal automation platform
Agent loops that self-certify their own success lie to themselves by default — a claimed_success report is optimistic, not verified. And loops that retry blindly on failure waste budget on objectives the strategy that just failed will never solve, because they never actually try anything different. Both failure modes look fine in the logs right up until someone checks the real acceptance criteria.
How it works
Three steps. No faith required.
01
Execute
A strategy runs against an objective and reports a result — but claimed_success is never trusted on its own.
02
Verify
A separate function recomputes pass/fail purely from acceptance_criteria, independent of what the executor claims.
03
Switch & seal
A failed objective gets a genuinely different strategy next attempt, not a repeat; recovery rate and final success both have to clear their floors before the packet seals.
Proof band
One packet. Every claim accounted for.
packet rig/operator · plan-execute-verify ops loop claim "recovery rate ≥70% on first-attempt failures, final success ≥85%" fixture 20 objectives · 5 traps (1 structurally unrecoverable) · seed 42 recovery PASS — 4/5 = 80.0% ≥ 70% floor final_success PASS — 19/20 = 95.0% ≥ 85% floor cross-check PASS — builder metrics match independent recompute themis SEALED — terminal verifier signed
Disable strategy-switching and all 5 trap objectives fail every attempt — recovery drops to exactly 0/5 = 0.0%, gate goes RED, nothing seals. The test then restores switching and confirms the gate returns to GREEN. How ProofPackets work →
Before you install
The obvious objections
Q Isn't "retry until it works" just... retrying?
Blind retry burns budget on a strategy that already failed. operator's select_strategy() cycles to a genuinely different approach on each retry — and verify() ignores the executor's own claimed_success entirely, recomputing pass/fail from acceptance_criteria alone. Nothing self-certifies.
Q How do you know the 80% recovery number isn't a friendly benchmark?
25% of the 20 objectives are deliberately traps the default strategy can't clear — and one is structurally unrecoverable by design, so 100% recovery was never on the table. 4 of 5 traps recover via strategy-switch; the fifth stays failed on purpose.
Q What if the executor lies about success and the verifier trusts it anyway?
It can't — verify() recomputes from acceptance_criteria minus criteria_met, a separate function that never reads claimed_success. Disable strategy-switching and recovery drops to exactly 0/5 — a hard, reproducible RED, not a soft miss.
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
20 objectives a month — the same 20-objective fixture the gate itself tests. Public evidence.
Team
Priced after a short conversation
Unlimited objectives, custom strategy repertoires, private acceptance criteria.
Pro
Priced after a short conversation
Dedicated verifier, CI-gated recovery regression, refute-turnaround SLA.
Runs well with
