Justin Leopard & how I think about agent infrastructure.

Shape

Software engineer, ten-plus years. The last two have been spent building agent systems that survive contact with real-world work — the kind where cost, trajectory, and escalation logic matter more than a clever prompt.

I read the papers the week they drop, then ship them as infrastructure the week after. Most of them don't survive the translation. The ones that do become defaults in whatever I'm building next.

What I build

The plumbing: orchestrators, memory stores, review gates, trajectory replay, cost-aware routing. The unglamorous layer between a planner and the humans who have to explain its decisions. If an agent fleet is going to learn across runs instead of within a prompt, this is the surface area that makes it possible.

The public substrate is the *-mini family: safe-mini, lab-mini, route-mini, and memory-mini as minimal, MIT, well-tested references for safety, labbing, routing, and durable memory.

What I believe, currently

• Handoff is load-bearing. Prompts are the glamorous layer; handoff and memory are the ones that decide whether a run compounds or collapses.
• Trajectories beat vibes. Evaluation only means something if the next run can retrieve and argue with the previous one's failure mode.
• Routing is a policy. Escalation should be derived from trajectory history — what's worked, what's burned budget — rather than whichever model the team happens to trust this week.
• Cost matters and is legible. Any observability story that doesn't render cost-per-outcome isn't one.

How I work

Local-first where it matters. Cloud where it earns its keep. Public writing as a forcing function — if a system's shape can't be explained on a page, it's usually too complicated to be right. I do not confuse ambition with disorder: build as ambitiously as the work deserves, but respect dependency order, test the claims, and keep security in the foundation instead of bolting it on after the launch.

Talk to me about

• Multi-agent orchestration and review gates.
• Typed memory and trajectory retrieval.
• Cost-aware routing and escalation policy.
• Evaluation layers that aren't just a scorecard.
• "The agent keeps doing the same expensive thing."