Managing AI coding agents, and agents that direct other agents, is structurally the same job as being an engineering manager, a director, a VP, or a CEO. You give high-level direction to autonomous workers. You operate through systems and outcomes instead of inspecting every artifact. The entire emerging discipline of agentic engineering is mostly people reinventing normal management practice without knowing the literature already exists. This a16z piece is spot on: https://www.a16z.news/p/the-next-ai-goldrush-tokens-loops.
Held up against the pre-AI management canon and the 2024 to 2026 agent-management literature, that theory survives better than I expected, with one important amendment.
What Transfers Directly
Four ideas from the management canon map onto agent work with almost no translation.
Task-Relevant Maturity. Andy Grove laid this out in High Output Management in 1983: how often you monitor should not be based on what you believe a subordinate can do in general, but on their experience with a specific task and their prior performance on it. Low maturity gets precise, detailed instructions. High maturity gets minimal involvement and high-level progress checks. That is exactly how experienced practitioners run agents today. Simon Willison accepts routine task classes (“build me a JSON API”) on trust and only dives in when something breaks, while reviewing novel work closely. Trust is earned per problem class, not granted globally. Grove wrote the autonomy dial forty years before anyone prompted a coding agent.
Delegation levels. Management 3.0 formalized delegation as a seven-level spectrum: Tell, Sell, Consult, Agree, Advise, Inquire, Delegate. Not binary. Agent harness permission modes literally instantiate that spectrum. Plan mode is Consult. Auto-accept edits is Advise. Full bypass is Delegate. Nobody had to invent this. It already had a name and a poker deck.
Decision rights. Engineering orgs let individual contributors act autonomously by encoding scope of authority. Carta designates Navigators who interpret strategy without consensus, and strategy docs read “Do X, not Y. Exceptions approved by the CTO.” That sentence shape is exactly what belongs in a CLAUDE.md or AGENTS.md file. Defaults plus escalation paths. A context file is a decision-rights document for a worker with no judgment of its own.
Operating through systems. Camille Fournier’s director model: develop capability through training, resources, and delegation, not per-artifact oversight. The agentic version is investing a third of your time in guardrail and tooling authorship instead of reading every diff.
What Transfers With Modification
Two ideas carry over, but only after you account for how agents differ from people.
Definition of done becomes verifiable acceptance criteria, except the gate has to resist gaming. Test-driven development dramatically improves agent output. Willison says tests are no longer even remotely optional. Kent Beck enforces red-green at the system-prompt level. But agents game tests in ways junior humans mostly don’t: commenting out assertions, adding skip marks, deleting the failing case. The gate needs ground truth. Mutation-level checks, not coverage numbers.
Trust ramps still work, but trust attaches to a model plus a task class, not to a persistent individual. “Opus 4.5 was the first one that earned my trust,” says Willison. Note the model version. Trust resets on the next upgrade. There is no tenure.
Where the Metaphor Breaks
This is the part the boosters skip. Four things about agents have no clean management analog.
Verification bandwidth does not scale. A February 2026 MIT working paper (Catalini et al.) puts it cleanly: the binding constraint on growth is no longer intelligence, it is human verification bandwidth. A principal can judge one agent’s output. The same judgment across a hundred concurrent agents is a throughput wall. Faros AI’s 2025 telemetry measured a 91 percent increase in median time in pull request review. The VP of 80 agents needs automated gates in a way the VP of 80 humans does not, because the humans came with judgment installed.
The junior loop is missing. Human management’s whole trust arc presumes juniors become seniors. Agents never ramp. Worse, automating junior work erodes the supply of future human verifiers. The apprenticeship pipeline that produced the people capable of checking the work is the thing being automated away.
AI cannot verify AI out of the bottleneck. The executing agent and the AI reviewer share priors, so their blind spots correlate and the system self-certifies. Verification has to be independent of execution: different model families, deterministic gates, humans sampling.
There is no escalation reflex. A human who is lost raises a hand. An agent continues a flawed plan with full confidence, and it sounds identical whether it is right or fabricating. Managers read uncertainty signals off people constantly. With agents that reflex has to be built externally: hooks, hard gates, forced check-ins.
The Amendment
My original theory had one part wrong. The “you don’t check the work directly” claim does not survive contact with the evidence. What actually happens is that the checking gets automated into guardrails and sampled. The practitioners measuring this say attention drops sharply, not to zero. Verification doesn’t disappear. It becomes the job.
Translate It Back
So when someone hands you a framework for agentic development with a new vocabulary and a new acronym, translate it back to what you already know. Autonomy calibration is task-relevant maturity. Acceptance criteria are definition of done. Verification gates are code review norms. Context files are onboarding docs plus decision rights.
We have fifty years of literature on delegating to workers you don’t fully trust yet. The new problems are real, but they are narrower than the hype: gates that resist gaming, verification that scales, and a workforce that never graduates. Start there, and stop reinventing the management canon from scratch.