Blog

We set out to measure one LLM's hiring bias against a caregiving career gap. The honest answer: the same model, on the same candidates, reads as anywhere from 0% to 100% biased — decided almost entirely by how you measure. Framing, metric normalization, and the workflow wrapper each swing the number more than the model does. The effect replicates identically on two model families, and one sentence removes it. A field note on why a bias number without its method is noise.

We graded AutoAW's genome against the design principles in a recent book on building multi-agent systems. The orchestration patterns, memory strategies, agents-as-tools, routing, and cost-aware fitness a human expert would reach for are, to a striking degree, already heritable traits under selection pressure.

Production agents remember in two fundamentally different ways: the host silently curates the context window, or the agent calls remember/recall tools itself. Instead of picking one, AutoAW makes the choice a heritable trait and lets evolution decide per task.

MCP standardized tool access. A2A standardized agent communication. AWP is the missing piece: an open protocol for defining portable, framework-agnostic multi-agent workflows.

Should we chase a single agentic workflow that generalizes across every task, or maintain a pool of specialized architectures we retrieve from per task? Why a fast, graph-aware registry of workflows wins on accuracy, cost, and latency.