A hand-written rubric for a machine-written search
AutoAW evolves multi-agent workflows. It doesn't read books — it runs a co-evolutionary genetic search over topologies, prompts, tools, memory, and delegation patterns, and keeps whatever scores well on your benchmark. So we tried an experiment of a different kind: we took an expert's hand-written account of how to build these systems — Victor Dibia's Designing Multi-Agent Systems (2025), a from-first-principles tour of patterns, memory, evaluation, and the failure modes that sink agents in production — and used it as an external rubric. If a human expert lists the levers that matter, does our search-space actually contain them?
The result was reassuring: the genome already encodes most of the book's design vocabulary. Evolution didn't need to read the playbook — the playbook is, to a striking degree, already in the search space. Here is where the two line up.
What the genome already gets right
The book's taxonomy of orchestration patterns, agent building blocks, and memory strategies maps almost one-to-one onto fields a gene already carries. Evolution didn't need to read the playbook; the playbook is mostly already in the search space.
Workflow vs. autonomous orchestration
The book's core spectrum — explicit, developer-defined control versus emergent, model-driven control — is our topology field (fixed_pipeline vs. ai_orchestrated), and the named coordination shapes (plan-based, fan-in, pipeline, best-of) are our delegation_pattern.
Managed vs. agentic memory
The book draws a sharp line between memory the host manages and memory the agent manages with tools. That's exactly the five values on our agent.memory slot — and we've written about why we let evolution pick.
Agents-as-tools, handoffs, and MCP
The “an agent is just another tool” composition pattern is our worker_mode: advisor | subagent; the handoff/router pattern is a per-turn router agent with router_mode; standardized tool access is mcp_servers.
Cost-aware delegation
The book treats cost and latency as first-class UX constraints, not afterthoughts. Our fitness is already a weighted blend of quality, cost, and latency, so an expensive workflow has to earn its tokens.
The headline: the book's design levers are, to a striking degree, already heritable traits — under selection pressure, competing on real benchmarks. The patterns a human expert would reach for are the patterns the population can already reach.
The genome already knows the playbook
It's a strange kind of validation to hold a machine-written search up against a human-written book and find they mostly agree on what matters. The orchestration spectrum, the delegation shapes, the two philosophies of memory, agents-as-tools, routing, standardized tool access, cost-aware fitness — none of these had to be bolted on. They were already heritable traits, already under selection pressure, already competing on real benchmarks.
That's the quiet promise of architecture search: you don't have to know in advance which pattern a task wants. You encode the vocabulary once and let the population discover the right dialect, task by task. The playbook describes the space; evolution explores it.
Building something adjacent?
Tell us which design patterns your task distribution keeps reaching for — that's the conversation we keep wanting to have.