ADR-153 — Darwin Mode: mutate one of seven harness policy surfaces (planner/contextBuilder/reviewer/retryPolicy/toolPolicy/memoryPolicy/scorePolicy), sandbox-score each variant, promote only measured wins. The WRITE layer that closes the loop ADR-150 opens (score+genome describe; evolve changes)....
AI agents invoke metaharness_evolve to trigger actions in Ruflo. What it does depends on the arguments the agent supplies, and its effects often reach beyond the immediate call — builds kicked off, notifications sent, workflows started.
This tool mutates core harness policies (planner, contextBuilder, reviewer, retryPolicy, toolPolicy, memoryPolicy, scorePolicy) and promotes variants that score better — effectively self-modifying the AI agent's governing policies.
From the tool's definition mutate one of seven harness policy surfaces, sandbox-score each variant, promote only measured wins... closes the loop... evolve changes
Attacks that exploit this kind of access
ADR-153 — Darwin Mode: mutate one of seven harness policy surfaces (planner/contextBuilder/reviewer/retryPolicy/toolPolicy/memoryPolicy/scorePolicy), sandbox-score each variant, promote only measured wins. The WRITE layer that closes the loop ADR-150 opens (score+genome describe; evolve changes). Use when readiness scores are flat and you want to discover WHICH surface mutation moves them, without retraining the foundation model. Bypassing this tool and hand-tuning is wrong because (a) single-degree-of-freedom mutations keep causal attribution clean, (b) the upstream safety layer catches secret/shell-out/network/dynamic-eval patterns before any variant runs (exit 99 = safety-disqualified, propagated verbatim). REQUIRES --confirm; defaults to dry-run plan output. Long-running: timeout scales with generations×children×sandbox-cost. It is categorised as a Execute tool in the Ruflo MCP Server, which means it can trigger actions or run processes. Use rate limits and argument validation.
Register the Ruflo MCP server in PolicyLayer and add a rule for metaharness_evolve: allow, deny, rate-limit, or require approval. Point your MCP client at the PolicyLayer proxy URL and the rule is enforced on every call, before it reaches Ruflo. Nothing to install.
metaharness_evolve is a Execute tool with high risk. Execute tools should be rate-limited and have argument validation enabled.
Yes. Add a rate_limit block to the metaharness_evolve rule in your PolicyLayer policy. For example, setting max: 10 and window: 60 limits the tool to 10 calls per minute. Rate limits are tracked per agent session and reset automatically.
Set action: deny in the PolicyLayer policy for metaharness_evolve. The AI agent will receive a policy violation error and cannot call the tool. You can also include a reason field to explain why the tool is blocked.
metaharness_evolve is provided by the Ruflo MCP server (ruvnet/ruflo). PolicyLayer sits as a proxy in front of this server to enforce policies before tool calls reach the server.
metaharness_evolve is one line of Ruflo's registry record.
The record carries the whole server: verified identity, auth posture, risk grade, every tool classified, recommended policy — re-checked continuously.
Teams ship this data inside their own products. See what a licence covers →