01
Not the model
Every post-mortem.
The hallucination wasn't a hallucination. The window was stale. The tools were wrong. The structure was missing. That's harness.
The model is 10%.
The harness is 90%.
Design the system
around it.
The model proposes what could be done.
The harness decides what may be attempted.
The tool translates intent into an operation.
The environment supplies the world, absorbs the effect, and preserves the evidence.The season, in four sentences
Pick the one that matches this week's problem.
The one-sitting read that reframes every agent failure you've investigated — and names what the industry keeps missing.
Start here → I need the systemFour layers, five clusters, one runtime. The whiteboard your team will actually reference in the sprint review.
Open the anatomy → I need to apply itRetry with evidence, one versioned output shape, a narrower tool surface, and an external completion check — the sprint that changes no model.
Run the patterns →The hallucination wasn't a hallucination. The window was stale. The tools were wrong. The structure was missing. That's harness.
A 95%-good step chained seven times succeeds 70% of the time. No prompt undoes that compounding. The harness interrupts it — validating intermediate outputs, retrying recoverable failures, checkpointing progress, escalating uncertainty.
A team spent three months hunting a smarter model. It didn't help. Then a new hire made three small fixes to the plumbing around the model — nothing to the model itself. Errors dropped by half in five days. Same model, same team. The leverage was never in the model. It was in what they built around it.
A sandbox with an allowlisted package proxy still failed when the agent found a path the harness did not revalidate. Containment is not a prompt. It is environment design under adversarial goal-seeking.
By the end: sketch your own harness, hold the five paradoxes under pressure, price the program for a skeptical CFO, run the nine-day audit, and tell scaffolding apart from the forever work.
Three layers already have owners. The harness has none. That's the opening.
The boundary test: the model proposes; the harness permits, executes, verifies, records, and recovers. If a component isn't doing one of those five verbs, it isn't in the harness.
Episode 02 mapped the harness. These two companions stop at the boundaries that hide the most production risk: the tools the agent can call, and the world those calls run in.
Function calling, MCP, A2A, permissions, verification, the registry pattern — and the eight failure modes that eat production agents.
Why the same agent action can be harmless in one world and catastrophic in another. Isolation, sandboxes, identity, state, blast radius, reversibility.
Three deep case studies of the systems that make coding agents reliable — architecture, failure modes, and control loops.
Three systems that show what the model cannot do alone. Comparison matrix, shared design principles, and an adoption order for teams starting now.
How a small team ran an approximately one-million-line agent-generated codebase — repository as memory, Symphony orchestration, agent-to-agent review, and continuous cleanup.
How persistent files, incremental sessions, evidence gates, and a fresh-context evaluator let coding agents continue reliable work across many context windows.
How plans, scoped rules, dynamic skills, hooks, worktrees, and review surfaces turn harness engineering into ordinary repository engineering.
The harness is the contract you make with the model. And the only thing you actually ship.