Agent Products Need a Control Plane and a Work Plane

The strongest July 10 signal clustered around OpenAI’s GPT-5.6 release and the launch of ChatGPT Work, which Digg elevated immediately as a top product story. The official release notes make the product direction explicit: Work is for longer tasks that can move across apps, files, and scheduled runs, while GPT-5.6 adds stronger computer use, artifact generation, and parallel workstreams. Simon Willison’s write-up sharpens the practical takeaway for builders: the interesting part is not just a stronger model, but a stack that can delegate, steer, and complete bounded work more efficiently. The market signal is that “chat” and “agent work” are no longer the same surface, even when they live in the same product.

Many teams still treat agents like an extended conversation that occasionally writes code or drafts copy. That creates muddled expectations: users do not know when they are brainstorming, when the system is actually acting, what it is allowed to touch, or what proof will come back. The next generation of useful AI products will make those states legible. The control plane holds goals, context, approvals, and redirection. The work plane holds execution, tools, budget burn, verifiers, and artifacts. That separation is what makes autonomy feel useful instead of vague.

Redesign one agent workflow as a two-plane product loop. Trigger: a user asks for a report, code change, launch asset, research brief, or other output that should take more than one model turn. Context: objective, source material, constraints, quality bar, and permissions live in the control plane. Tools: the work plane gets the model route, app connectors, browser or file tools, and any subagents needed for bounded execution. Verifier: require a visible acceptance test such as passing checks, source-backed claims, screenshots, or a reviewable artifact before the run can claim completion. Budget: define time, token, tool, and privilege limits up front, plus which actions require approval. Artifacts: return a concise worklog, the finished deliverable, evidence of verification, and the next decision the human needs to make. Stop condition: the run either produces a verified artifact inside budget or pauses with a named blocker and a clear request for more context or permission.

Take one agent feature you ship today and mark every moment where the user is thinking versus every moment where the system is acting. If those states blur together, add explicit handoff, budget, and verification steps before you add more autonomy.