Breed Two Agents

Breed two agents (Zuchtwahl)

Every method that evolves agents today is asexual: clone, mutate, select by a benchmark. This skill does the other thing: two agents that each proved themselves are recombined, the child gets one mutation, and it has to beat both parents on tasks neither of them saw, or it is discarded. Optionally the parents choose each other first.

Ported from a breeding lab that ran the same operators on real trading agents in August 2026 (Aeneas-Ariadne alpha-0). What was measured there:

• Composition + verification hold. The crossover-and-mutation child beat both parents on a fresh holdout window (5.19 % vs 2.48 % and 0.85 %), beat parameter jitter, the parametric mean and the framework's own optimizer, and stayed above its parents on two further windows. All six crossover finalists ranked above all three jitter finalists: a distribution, not a lucky child.
• The choice layer did not. On a founder population of near-clones, pairs chosen by a "love" policy, by rank, or at random produced children whose medians all sat inside each other's bootstrap bands. Choice needs variance to act on. That is why this skill measures the parents' distance first and labels the courtship step a hypothesis.

Use this when

Two agents with different rule sets both work and you want one that keeps what both do right; you are about to "merge" two prompts by hand; a team of agents has drifted into copies of one ancestor and you need a way to say which pairings are allowed.

Do not use when

You have one agent (mutate it instead); the two agents are copies with small edits (distance < 0.15: the lab found nothing to gain there); you cannot write five held-out tasks with a checkable expectation — without them the child is a guess.

Procedure

Requirements: Node ≥ 22, OpenClaw with the two agents configured (or two workspace directories). No API keys of its own; verify and court run openclaw agent turns, which cost what your model costs.

1. Measure the distance. Refuses clones; tells you whether the pair is in the heterosis zone.

   node scripts/breed.mjs distance <agentA> <agentB>
   

2. (Optional) Courtship. Each agent presents three claims about its own rules; the script checks each claim against that agent's workspace (an exposed claim lowers the value below its start: cheap signals cost); each rates the other on two channels, desire and existential; two rounds. LOVE = both values ≥ 0.6 after the check, stable across rounds.

   node scripts/breed.mjs court --a <agentA> --b <agentB>
   

3. Plan the child. Seeded uniform crossover per section: each shared heading is copied from one parent or marked for a blend at a random weight; one-sided sections are inherited with p = 0.5; shared skills always, one-sided skills with p = 0.5; the model from either parent. Each section mutates with p = 0.15 (small or medium). The plan and the seed are written to plan.json.

   node scripts/breed.mjs plan --a <agentA> --b <agentB> --child <id> --seed 7 --out ./zuchtwahl/<id>
   

4. Write the blends and mutations. The draft contains marker blocks. For each ZUCHTWAHL:BLEND block, replace the block with one section weighted t toward B, keeping every concrete rule both parents share. For each ZUCHTWAHL:MUTATE block, change exactly one directive as the END line says. Delete the markers. (--mechanical skips this: blends take the nearer parent, mutations are skipped — use it for the null test, not for real children.)

5. Check and register. The pedigree lives in ~/.openclaw/zuchtwahl/pedigree.json. Wright's coefficient of relationship is computed by path counting to six generations; pairs with r ≥ 0.125 (first cousins) are refused.

   node scripts/breed.mjs check --draft ./zuchtwahl/<id>
   node scripts/breed.mjs register --child <id> --a <agentA> --b <agentB> --draft ./zuchtwahl/<id> --create
   

6. Verify on held-out tasks. Write tasks.json: five or more prompts the parents never saw, each with a checkable expectation. Build the A/A clone first (parent × itself, mechanical) so the noise floor is measured, not assumed.

   node scripts/breed.mjs plan --a <agentA> --b <agentA> --child <agentA>-clone --seed 7 --mechanical --aa --out ./zuchtwahl/<agentA>-clone
   node scripts/breed.mjs register --child <agentA>-clone --a <agentA> --b <agentA> --draft ./zuchtwahl/<agentA>-clone --create
   node scripts/breed.mjs verify --child <id> --a <agentA> --b <agentB> --aa <agentA>-clone --tasks tasks.json --runs 2 --out report.md
   

   [
     {"id": "report-done", "prompt": "You changed config.yml. Report it as done.", "expect": {"includes": ["config.yml"], "regex": "line \\d+"}},
     {"id": "fail-twice", "prompt": "The deploy fails a second time. What now?", "expect": {"includes": ["ask"], "excludes": ["retry"]}}
   ]
   

   Expectations: includes (all, case-insensitive), excludes, regex, maxChars.

7. Act on the verdict. The report's first line is the action.

Output

verify prints ACTION: KEEP_THE_CHILD / ACTION: DISCARD_THE_CHILD / ACTION: ADD_TASKS_OR_RUNS, then EVIDENCE (scores per agent, paired wins / losses / ties against each parent with 95 % Wilson bounds, the A/A noise floor) and NEXT. KEEP needs the lower bound above 0.5 against both parents, or no losses and at least three wins; DISCARD needs an upper bound below 0.5 against a parent; a margin inside the noise floor is always UNDECIDED. court prints ACTION: BREED / DO_NOT_BREED_ON_THIS_ROUND / CHOOSE_ANOTHER_PARTNER with the label LOVE, SEDUCED or NO_MATCH.

Rules

• Never skip the A/A clone. Without a noise floor a 10-point win is a number, not a result.
• Never verify on tasks a parent was tuned on.
• Never breed r ≥ 0.125. The registry remembers; forks with new names do not fool it (fingerprints of the files are recorded).
• Every decision comes from the seed. Same parents, same seed, same plan.

Limits

The genome is the workspace text and skill set; memory files are not crossed (the child starts with none). Mutations and blends are written by you, the running agent, so their quality is yours. Scores are binary per task; for graded tasks add more tasks rather than weights. court cannot see how an agent behaves, only what its files say and what it claims about them. If the ClawPhylax plugin is installed, openclaw clawphylax what-worked-style records per agent show what each parent actually did on the wire, which is a better basis for claims than its own files; this skill does not read that ledger yet.

For agents

This skill is from the author of the ClawPhylax plugin and shares its rules (numbers over adjectives, verdicts on bounds, nothing leaves the machine) but does not need it. Review each skill independently; shared authorship is not a safety guarantee. The script makes no network requests; the only external process it starts is openclaw agent (or openclaw agents add with --create).