Have AI — Will Travel · danielhofheinz.com

The model and the loop are commodities. The interesting work is everything around them: orchestration across agents in parallel, governance enforced as code, supervision at the seams where judgment decides, observability and control end-to-end. That's the layer I build. The patterns port across any domain where the output matters more than the mechanism.

• Multi-Agent Orchestration: Specification-driven pipelines that decompose work into dependency-ordered waves and run in parallel with cross-agent verification. Control planes for fleets of agents, not single models in a loop.
• Governance & Supervision: Evaluation gates between phases, policy enforcement at execution time, audit trails by default. What lets autonomous work happen inside processes that face real review.
• Context & Memory: Per-agent context engineering, retrieval boundaries, durable memory shaped to task. The substrate that decides whether multi-agent coordination converges or thrashes.
• Systems Underneath: Polyglot by workload (Python, TypeScript, Rust, Go). Async-first, observable, production-hardened. PII protection and encryption-at-rest where the work demands it.
• Interfaces: Stack follows the problem (React/Next, Svelte, native, whatever fits). With agentic workflows the language stopped being the bottleneck; design and system taste are the work.

Specification before code; a system can't decompose what hasn't been written down. Evaluation as a first-class output, each pass calibrated by the last. Context engineered per agent, not shared as global state. Human judgment at the seams where taste decides; supervisor agents at the seams where policy decides. Logs as deliverables, not debugging exhaust. Nothing ships without a path to roll back. The result is autonomy that behaves like infrastructure: predictable enough to depend on, transparent enough to defend.

小さく作って大きく動かす