2026-10-13
October 13, 2026
Every team running AI agents has already answered the question "who decides?" The answer is usually whatever the default permission prompt happened to be.
Ask a team how much autonomy their coding agents have and you will get a tooling answer: which product, which plan, whether auto-approve is on. Ask who decided that level, on what evidence, and when it was last revisited, and the room goes quiet. Autonomy is being granted one click at a time by individual engineers, and the sum of those clicks is your organization's real delegation policy.
This is a very old problem wearing new clothes. Managers have always had to decide how much of a decision to hand to someone else, and the good ones made that call explicitly, per decision, and adjusted it as trust grew. Jurgen Appelo built an entire practice around it, Delegation Poker, with seven levels running from "I decide and tell you" to "you decide, no need to tell me." Nobody would run a human team by letting each person silently pick their own level on every decision. We are doing exactly that with agents.
The senior-leader move this quarter is small and unglamorous: make the delegation level a named, reviewed property of each class of agent work, set by risk and reversibility rather than by enthusiasm. That is what today's issue is about. The vocabulary already exists; it just has not been applied.
Start with what is actually being decided. Addy Osmani's "Agentic Autonomy Levels" (July 2, 2026) splits autonomy into two axes: agency, meaning how far a single agent acts on its own, and orchestration, meaning how many agents run and who coordinates them. He lays out six levels, from Level 0 (the agent suggests, a human acts) through Level 2 (a bounded task with a clear definition of done), Level 3 (goal-driven loops of plan, act, test, review), Level 4 (parallel agents on isolated slices) and Level 5 (a manager agent that dispatches workers and escalates only exceptions). The useful claim is not the ladder. It is that the ceiling is set by risk and reversibility, and that the level should follow the strength of your verification, not the label on the task.
Osmani offers three questions for judging how much autonomy a system really has: how quickly would you detect that it was wrong, how cleanly could you undo its work, and what would prove it right? Read those as a manager would. They are the same questions you ask before delegating a decision to a human: what is the blast radius, how fast does feedback arrive, and how will I know it went well. A TPM who has ever written a rollout plan already owns this reasoning.
Dennis Pilarinos at Unblocked (June 22, 2026) puts it as a minimum function: autonomy is limited by the lower of the agent's context quality and the team's ability to catch wrong output. If the agent has thin context, more autonomy produces confident mistakes. If your review is weak, more autonomy produces mistakes nobody catches. His warning is the one I would put on a slide: the clean-looking pull request is the dangerous one.
Now apply the human-delegation lens. Appelo's seven levels are Tell, Sell, Consult, Agree, Advise, Inquire, Delegate. The game works because it forces a per-decision answer and surfaces disagreement: everyone reveals a card at once, the highest and lowest explain themselves, the group converges. The mapping to agents is my own adaptation, not Appelo's, but it holds up. "Consult" is an agent that drafts and a human who decides. "Agree" is a human approving each consequential step. "Advise" and "Inquire" are an agent that acts within a contract and reports afterward. "Delegate" is Osmani's Level 5, managed by exception. The point of borrowing the vocabulary is that your stakeholders already understand it.
What goes wrong without it? Osmani names four anti-patterns that read like a management audit. Autonomy as status: teams brag about the highest level instead of the right one. Permission laundering: broad access granted because narrow scoping was tedious. Summary substitution: reviewers read the agent's summary instead of the evidence. Fleet cosplay: many agents running while humans still coordinate dependencies by hand. Each is a governance failure, not a model failure, and each is detectable by a program lead who asks the right question in a review.
Finally, the cost dimension. A FinOps Foundation piece published October 6, 2026 argues that token and API charges are only a slice of what AI actually costs once infrastructure, tooling and labor are counted, and that organizations with a defined owner for AI economics were far likelier to show value to their CFO. The article attributes that to its own survey data, so treat the multiples as directional. The connection to delegation is direct: every step up in autonomy changes spend, review load and failure cost at once, so the level and the budget should be set in the same conversation by the same owner.
Put together, the practice is a delegation register: one row per class of agent work, with a named human owner, a delegation level, the verification evidence required, the reversibility rating, the budget, and a review date. It is short, it is auditable, and it turns an invisible default into a decision someone can be held to. Osmani adds a calibration ritual worth stealing: look at the last ten agent-assisted tasks and record the level used, the risk, the evidence, the review time, the rework, and whether the level was right. You will find that some work was over-supervised and some was dangerously under-supervised, usually on the same team.
Try this week. Pick one agent workflow your team already uses and run a ten-minute Delegation Poker on it. Have three people, an engineer, a reviewer and the on-call owner, privately choose a level for "merge without human review," reveal together, and have the highest and lowest cards explain themselves. Write down the agreed level, the evidence that justifies it, and a date to revisit. That is your first register row.
What it is. A team exercise from Management 3.0 in which participants use seven cards (Tell, Sell, Consult, Agree, Advise, Inquire, Delegate) to agree on how much authority a manager and a team each hold over a specific decision.
When to use it. When ownership of a recurring decision is fuzzy or contested: who approves a production deploy, who accepts an agent's output, who may change an on-call policy. It is especially useful when a new capability, such as an AI agent, shifts who could reasonably decide.
How to run it:
When NOT to use it. For decisions that are legally, security-wise or contractually fixed, where there is nothing to negotiate and the exercise only creates false ownership.
Example: a team discovers its engineers think agent-authored dependency upgrades are at "Inquire," while the security lead thinks they are at "Consult." That gap is the finding.
AI Spend Budget Busters: Expanding the Budget Planning Lens — Tammy Burnitt (FinOps Foundation, Oct 6) argues token cost understates total AI spend and recommends planning across pre-deployment, the first 30 to 90 days, and the steady state; the ownership point is the part to carry into your next budget review.
How to prepare for AI agents — Robert Kimani on LeadDev (Oct 2025) recommends an agent registry with owners, access and rollback plans, plus scoped credentials and cost caps before scaling; the register idea in today's issue extends it.
How Much Autonomy Should You Give Your AI Coding Agents? — Dennis Pilarinos (Unblocked, June 2026) frames autonomy as limited by context quality and your ability to catch bad output; it comes from a vendor with a product to sell, so use the framing rather than the pitch.
"Trust, but verify."
— Russian proverb, popularized in English by Ronald Reagan
Don't miss what's next. Subscribe to Critical Path: