Would anyone pay for the harness?

The phrase has drifted into a catch-all, so the first job was to draw lines. In practice the demand clusters into three shapes:

• Augmentation, an agent embedded in an existing process (triage, drafting, enrichment) that makes a human faster without replacing the decision.
• Orchestration, a multi-step pipeline where the agent owns the sequence: fan a job across sub-tasks, handle retries, hold state across steps, and refuse to finish until a gate passes.
• Autonomy with a leash, a workflow that runs unattended on a schedule but reports to a human gate before anything irreversible.

The interesting finding: most buyers ask for the third and actually want the first. They want leverage, not abdication.

The model is a commodity, everyone is one API call from the same intelligence. The defensible work is the harness: the verification, the state, the tool integrations, the human gates, the knowing which structure to remove. That maps almost exactly onto what's already running on this site, an enrichment orchestrator that fans across five sub-agents, a deterministic playtest gate that blocks a merge on a real oracle, a content pipeline that won't publish without approval. The research mostly confirmed a hunch: the sellable asset isn't "I can call an LLM," it's "I've shipped harnesses that fail safe in production, and here's the build log."

Research isn't a plan, and a few things stayed genuinely unresolved. Pricing is the big one: agentic work resists hourly billing because the leverage is non-linear, but fixed-scope offers need a verification boundary clear enough that "done" isn't a negotiation. Liability is the second, an autonomous workflow that's wrong unattended is a different risk class than a tool a human drives. The next step is small and concrete: take one real internal workflow here, the kind a client would recognize, and write it up as a worked case study with the gate, the failure modes, and the time saved, turning the research into evidence instead of a claim.