The model is one part of the system
An agent task needs somewhere to hold its conversation and state, a way to reach approved tools, and a way to stop or ask a person. A model can propose an action while the runtime still has to decide whether and how that action may run. The harness should make that distinction inspectable.
The Spinrun execution path
Spinrun's web app handles the product's request and account context. Agent work runs through agents-node and its Mastra-based runtime. Tool access passes through the gateway and workspace authorization. Saved history and live state serve different purposes: one preserves evidence, the other describes work in progress.
Authorization stays at the boundary
An agent instruction cannot grant membership or widen an external connection. The gateway checks the account and workspace scope for a tool call. Approval decisions apply to proposed actions in their context. Keeping these checks outside a model's reasoning prevents a fluent explanation from becoming a permission rule.
Recovery is part of correctness
An interrupted task may have a confirmed read, an unconfirmed write, and a saved draft. Treat those states differently. The caller should retry only when it knows the action could not have run, or when the operation is safe to repeat under a known idempotency rule. Inspect an uncertain destination before repeating a write.
Design for a human reviewer
Show the source and proposed action at the point where a person decides. A task's finish line should be an output the team can inspect. The harness adds value when it makes progress, assumptions, and failures understandable without requiring the reviewer to reconstruct every internal step.
Keep experiments bounded
Evaluate changes against representative tasks with a written rubric. Track useful output, failure behavior, review effort, and cost. A harness improvement should make a supported workflow more reliable; a general intelligence label is not a substitute for that evidence.