Harness evolution agent
Hill-climbed configs.
Cost-bounded.
rig-evolver mutates your coding-agent harness config — retry limit, temperature, max steps — across generations, keeping a mutation only when it beats the current best AND doesn’t regress cost or failure rate beyond threshold. The verifier re-simulates the held-out comparison from scratch.
rig-evolver · population hill-climb · elitist selection[higgsfield: gold-lit ascending staircase of glowing config knobs, dark workshop, shallow depth of field, brass dials catching key light]
Why teams install
What rig-evolver earns back
Wins on held-out tasks, not train-set luck
Initial vs. final config replayed on the same seeded eval set — 77.5% win rate at seed 42, 40 tasks.
‹9995ea5›Cost can’t creep to win
A mutation is discarded the moment cost rises >15% or failure rate rises >8pp versus the current best.
‹bd7ee1b›Verifier never trusts the builder’s own numbers
Given only the initial config, final config, and seed, the verifier re-runs the held-out comparison independently.
‹6690be2›How it works
Three steps. No faith required.
01
Simulate
Run the current config across a batch of tasks; every step attempt costs money and consumes budget.
02
Mutate
Bounded perturbations to retry_limit, temperature, max_steps drawn from the current best.
03
Compare
Paired, same-seed replay of initial vs. final config on held-out tasks — a real win, not a lucky roll.
Proof band
One packet. Every claim accounted for.
packet rig/rig-evolver · evolution loop v1 run seed 42 · 40 held-out tasks · population hill-climb result win_rate 0.775 (≥ 0.60) · cost_delta −17.3% (≤ +10%) config initial{t=1.4, steps=4} → final{t=0.329, steps=8} anchor PASS — verifier independently re-simulated the comparison ‹9995ea5› themis SEALED — merge gate armed
Every rig-evolver run emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Buyers
Who runs rig-evolver
Platform Eng Lead
agent-tooling startup
ML Infra Manager
enterprise copilot team
Founding Engineer
harness-heavy dev tool
Objections
Answered before you ask
Isn’t this just hyperparameter search?
Hyperparameter search optimizes one metric. rig-evolver rejects any mutation that regresses cost OR failure rate beyond threshold — a win that quietly gets more expensive is discarded, not accepted.
Could the evolved config just be overfit to the training tasks?
win_rate is computed on a held-out task set the verifier re-simulates independently — same seed, same tasks, but the verifier never sees the builder’s self-reported comparison.
What happens if evolution just does nothing?
Disable the loop and final_config==initial_config — the paired comparison shows exactly 0% win rate and the gate goes RED. That’s a regression test run on every build.
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
1 config, 20 held-out tasks a month.
Team
Priced after a short conversation
Unlimited configs, full generation history, cost dashboards.
Pro
Priced after a short conversation
Dedicated evolution fleet, per-team configs, SLA.
SOC2 Type II readiness targeted Q4 2026 — every generation is audit-exportable from day one.
Runs well with
