Volume IV · The model is the engine. The harness is the vehicle.

The Harness

Frontier models have converged, so the tools, permissions, and execution environment wrapped around them now decide the experience — this is also the layer where Operon lives. The volume runs from the agent loop and tool design through permission architecture, sandboxing, hooks and checkpoints, skills and subagents, to an evaluation rubric for choosing or building a harness.

Chapters

Published

01The Harness Decides the ExperienceModel convergence and the 17-harness/one-model benchmark result: why identical models feel different across tools.8 min readPublished

02The Agent LoopGather context, act, verify, repeat — the Ralph loop lineage and the "every harness is a loop" debate, addressed rather than dismissed.8 min readPublished

03Tools Are APIs for AgentsTool design as the new interface design: naming, granularity, error surfaces, token cost of tool results.9 min readPublished

04Permission ArchitectureThe tiered consensus: auto-approve read-only, gate writes and dangerous commands, feature branches never main.9 min readPublished

05Sandboxing & Blast RadiusOS sandboxes, container isolation, network egress control, worktree isolation — blast radius as a design budget.8 min readPublished

06Hooks, Gates & CheckpointsDeterministic lifecycle control points and checkpoint/rollback mechanics — every team convention living in a senior engineer's head can be encoded as enforced harness logic.8 min readPublished

07Skills & Instruction PacksThe skills primitive and the isolation spectrum — skills, subagents, agent teams — as progressive disclosure of instructions.7 min readPublished

08Memory & State in the HarnessMemory files, resumable state, durable checkpointing, crash recovery, rainbow deployments for running agents.8 min readPublished

09Long-Running & Background AgentsInitializer + incremental-agent patterns across context windows; overnight loops; the iterate-deeper vs. scale-wider debate.8 min readPublished

10Harness EconomicsToken spend as a function of harness design — plan-heavy, read-more-files harnesses vs. lean ones, and the severalfold same-task cost spread that pays for it.7 min readPublished

11Designing HarnessesCapstone — an evaluation rubric for choosing or building a harness.5 min readPublished

Key sources: Anthropic glossary + "Effective harnesses for long-running agents" + "Writing effective tools for agents" · AIMultiple 17-harness benchmark · Huntley Ralph posts · practitioner analyses of skills/subagents/hooks composition

Related