2026-10-27
October 27, 2026
Most AI governance answers "which tools can the agent touch?" The question that actually scales is "who decides, for this kind of decision?"
Look at how most organizations govern coding and operations agents today. There is a permissions list: repositories, credentials, environments. There is a policy document. There is, somewhere, a human who approves things. What is usually missing is the one artifact every TPM already knows how to produce for people: an explicit statement of who has the decision, and how much of it.
Access control tells you what an agent can do. It says nothing about what it should decide alone, what it should propose, and what it should only report. Those are different questions, and conflating them produces the two failure modes leaders keep describing: engineers who approve every step until they are exhausted, and engineers who quietly stop looking.
Management has had a vocabulary for this for years. This issue borrows it, applies it to agents, and argues that the delegation level, not the permission set, is the unit a senior TPM should be negotiating.
Start with the supervision numbers, because they explain the mood. Allstacks, summarizing Anthropic's 2026 agentic coding trends report, puts the pattern this way: developers use AI in roughly 60% of their work but can fully delegate only 0 to 20% of tasks, and they actively supervise 80 to 100% of what they do hand off. Treat those as a vendor's secondhand summary, not a measurement you should quote to your CFO. But they match what most of us see: the agent is fast, and the human is still the bottleneck, because nobody has said which outputs deserve scrutiny.
Fabien Potencier of Upsun frames the fix as a dial rather than a switch. Start with every action behind a human gate. After each run, do the retrospective: feedback, prompt changes, refined skills. As evidence accumulates, remove gates and move the steps that can be deterministic into scripts. The human steps back not because they were replaced but because the guardrails now show when their attention is needed. That is a good description of the trajectory. It lacks a vocabulary for where each decision currently sits on the dial.
Jurgen Appelo's Management 3.0 supplies one. His seven levels of delegation run from Tell (the manager decides and informs), through Sell, Consult, Agree, Advise, and Inquire (the team decides, then informs the manager), to Delegate (the team decides with full autonomy). Applied to an agent, the levels become operational. At Tell, the agent does nothing without instruction. At Consult, it proposes and a human decides. At Agree, it drafts and a human must co-sign before anything lands. At Advise and Inquire, the agent acts and the human gets a notification, either before or after. At Delegate, the agent acts and nobody is told unless something breaks.
The important move is that the level attaches to a decision class, not to a tool or a team. The same coding agent can sit at Inquire for dependency bumps with green tests, at Agree for schema migrations, and at Consult for anything touching billing. Seen this way, the exhausted-approver problem is a mis-leveled decision: a Consult-level gate on something that belongs at Inquire. The quiet-abandonment problem is the reverse: a Delegate-level behavior on something that deserved Agree, reached by default because nobody wrote it down.
Three things make this work in practice. First, verifiability sets the ceiling. Anthropic's report, as summarized by Allstacks, notes that agents do best on tasks that are verifiable, well defined, or repetitive, and fail expensively when inputs are ambiguous. If you cannot cheaply check the output, the decision class cannot go above Agree, whatever the agent's track record. Second, reversibility sets the floor for notice. A reversible action can sit at Inquire; an irreversible one needs a human before it happens. Third, levels move on evidence, in both directions. The LeadDev panel on agent autonomy, published in July 2026, lists the same two questions as its agenda: what conditions allow autonomy to expand, and what should trigger rolling trust back. Write the rollback trigger down when you grant the level, while nobody is angry.
This is also where the TPM earns the seat. Someone has to keep the register of decision classes, run the calibration conversations, and notice when the org has drifted into de facto Delegate on something nobody agreed to. That is program work: it is ownership, escalation paths and review cadence, applied to a new kind of team member. Aadil Maan has argued that TPMs should ask whether AI can "play the role" of a TPM across process, synthesis, distribution, accountability and troubleshooting. A delegation register is the accountability piece of that list, and it is the part least likely to be automated away, because it is a negotiation among humans about risk.
Try this week. Pick one agent your teams already use. List its ten most common actions, and write a delegation level next to each (Tell through Delegate). Then ask the owning engineer to do the same blind. Wherever your two levels differ by two or more steps, you have found either a wasted approval or an unexamined autonomy. Fix the three largest gaps.
What it is. A team exercise from Appelo's Management 3.0 in which participants privately pick one of seven delegation levels for a specific decision, reveal together, and discuss the differences.
When to use it. When a decision class has unclear ownership, when you are about to grant an agent more autonomy, or when a team and its manager keep disagreeing about who "should have" decided something.
How to run it:
When NOT to use it. For decisions already fixed by regulation or security policy, where the level is not up for debate; record those as constraints instead.
Variant for agents: have a human play the card the agent would effectively act on today, based on its configuration, and compare it to what the humans want.
Trust dial: how teams earn autonomy for AI agents — Fabien Potencier (Upsun, August 2026) argues trust must be built through shared context, skills and guardrails, with retrospectives after each run. Read it as the operating loop that sits underneath a delegation register.
The Orchestration Gap — Allstacks (April 2026), a vendor post that summarizes Anthropic's agentic coding report and argues for oversight systems and upstream spec quality. Useful for the supervision figures, with the usual caveat about the source.
What Happens to TPMs When AI Starts Running the Program? — Aadil Maan on whether AI can play the TPM role, and which functions stay human. Written in 2025 and speculative, but its five-function rubric is worth arguing with.
"Trust, but verify."
— Russian proverb (doveryai, no proveryai), popularized in English by Ronald Reagan
Don't miss what's next. Subscribe to Critical Path: