Forked Agent Architecture — Production Parallel Execution Design

Parallel agents sound simple until they share state. One agent writes a file at the same time another reads it. Two agents update the same data structure. A merge resolves in the wrong order. The output looks fine. The contamination is structural. This SKILL.md is the fork-merge architecture that prevents all three.

---

⚡ What's Inside

• Fork initiation protocol The exact pre-fork checklist. What must be serialized, what must be isolated, and the agent ID assignment scheme that prevents namespace collisions during parallel execution

• Isolated memory scopes How forked agents get read access to parent context without write access, and the boundary enforcement that makes that guarantee durable rather than advisory

• Safe parameter passing The parameter types that are safe to pass to forked agents and the ones that are not. The unsafe list is longer than you expect

• Cache-sharing pattern How forked agents share expensive context (embeddings, retrieved memory, tool results) without duplicating it across forks. Cuts cost and latency significantly on context-heavy parallel workloads

• Deterministic merge protocol Ordered merge sequence, conflict resolution rules, and the artifact checksum validation that catches merge errors before they propagate

---

🏭 Proven in Production

This architecture came from running parallel extraction agents against large session contexts. The shared-nothing model was too slow. The naive shared model produced contamination. The cache-sharing pattern with isolated write scopes was the middle path — and it took several failed attempts to get the boundary rules right.

---

🆕 What's New in v1.1

• Cache-sharing cost model added
• Merge conflict escalation path documented for human-in-the-loop resolution

---

✅ Core Capabilities

• ✅ Fork initiation checklist — pre-fork serialization and agent ID assignment
• ✅ Isolated write scopes — parent context readable, not writable, from forks
• ✅ Safe parameter passing spec — exact list of safe and unsafe parameter types
• ✅ Cache-sharing pattern — shared reads, isolated writes, zero duplication cost
• ✅ Deterministic merge protocol — ordered sequence with artifact integrity validation
• ✅ Works with any OpenClaw multi-agent topology