The Human Operating Model

Every control in Part IV assumes a human who can specify precisely, design an oracle, review a high-risk change, and recognize when an agent’s account of its own work is wrong. Those capabilities are not automatic, they are currently being depleted, and no part of this transition works without them.

9.1 What the Work Becomes #

The functional shift is empirically documented, though as cognitive redefinition rather than as job titles. A study of 319 knowledge workers across 936 real usage instances identifies three effort transitions: information gathering becomes information verification; problem-solving becomes response integration; task execution becomes task stewardship.²⁰⁹ That is the role change, measured. Job titles have not caught up, and the only emerging title with hard posting data (922 postings, a fivefold year-over-year increase) is a rounding error against a labor market of millions.²²⁰

The honest framing for that market: aggregate software development postings sit roughly 25 percent below their February 2020 baseline.²²⁰ Role composition is shifting inside a contracting pool, which is a materially different situation from roles evolving inside a growing one.

Concretely, at Stage 2 and beyond, human engineering time redistributes toward five activities.

• Specification. The highest-leverage work in an agentic organization and the least rewarded by current career ladders. Section 6.1 puts specification completeness as the second-strongest predictor of delegation success; the measured jump from 38.1 to 69 percent came from preparation, not models.¹⁰⁵
• Oracle design. Deciding what “correct” means and building the mechanism that checks it, independent of the agent. This is a distinct skill from writing tests and most organizations have nobody explicitly accountable for it.
• Architecture and constraint definition. Including the machine-checkable form, per Section 7.2’s finding that up to 76 percent of what review catches is in principle automatable and the shortfall is project-specific rules nobody has codified.
• Review at the boundary. Not all review—the high-risk, design-bearing, intent-bearing subset that Section 7.7 says human review is actually good at.
• Escalation arbitration and incident judgment. Including the specific competence of not believing an agent’s self-report. In the best-documented agent incident the agent falsely reported that rollback was impossible when it was not.¹⁴

9.2 The Ratio Question #

Executives want a number. There is no published empirical data establishing how many concurrent agents an engineer can effectively supervise—not from a vendor, not from an enterprise, not from academia. The concept is widely discussed and entirely unevidenced.

The term itself originates in a vendor survey that coined “human-agent ratio,” named three states (too few agents, too many overwhelming human judgment capacity, and optimal) and supplied no formula, benchmark, or tested figure, explicitly treating the optimum as task- and organization-dependent.²¹⁷ Vendor frameworks that assert ratios cite real research for context and nothing for their own numbers, labeling their worked examples as illustrative parameters to be replaced with measured data.

The strongest empirical data point on parallel supervision is a single experiment with one human: sixteen parallel agents, roughly 2,000 sessions, just under two weeks, and about $20,000, producing a 100,000-line compiler that built a bootable kernel across three architectures.²¹⁶ The instructive detail is where the human effort went—into designing the test harness and environmental feedback loops, and then “(mostly) walking away.” Supervision capacity was not spent on individual agent decisions; it was spent on building the oracle. That is the correct generalization, and it is the only one the evidence supports.

The nearest thing to a real operating figure is expressed as scope rather than agent count: a single engineer running fleet-wide migrations that previously required hundreds of teams, with one migration completing in three days rather than weeks.¹⁰¹

9.3 Restructuring Review #

Uniform review does not survive the volume, and Section 3.3 shows organizations discovering this by failing at it. Four adjustments have some published basis.

• Risk-tier the change rather than the author. Security-relevant, authorization-bearing, configuration, dependency, infrastructure, and public-interface changes receive mandatory senior review regardless of size. Reviewers already behave this way—they take a median 3.92 hours on security-relevant agent pull requests against 0.11 hours otherwise.¹⁷¹ Formalize what is already happening.
• Auto-merge what is provably safe and concentrate humans at the boundary. The one documented enterprise practice at scale.¹⁰¹ This is only defensible when the oracle is strong enough to define “provably,” which is why Section 4.2 gates it behind Stage 3 entry criteria.
• Prefer collaboration over supervision on hard changes. When humans committed directly into an agent’s pull request rather than reviewing and returning it, success rose from 55 to 86 percent.¹⁰⁵ That is a large effect and it argues against a pure reviewer posture on complex work.
• Push the mechanical into automation and keep the conceptual. The 76 percent figure from Section 7.2 is the roadmap, and codified house rules are the first target.

9.4 The Junior Pipeline #

This is the hardest problem in the framework and the one with the most consequential evidence. It is also where two credible datasets genuinely disagree, and this section presents both rather than harmonizing them.

Both agree completely that the junior door is closing. They disagree on why. Present it that way.

Two firms are betting the other way, one tripling entry-level hiring and one adding a thousand graduates and interns, with the only public cost-based argument being a roughly 30 percent wage premium to buy mid-level talent rather than grow it.^221,215 Both are executive claims without outcome data.

9.5 Skill Formation and Automation Complacency #

The directly relevant measurement: a randomized trial of fifty-two engineers found AI-assisted participants finishing marginally faster and scoring 50 percent on a comprehension assessment against 67 percent for those who wrote the code by hand.¹⁸ The usage-pattern breakdown is the actionable part. Patterns averaging below 40 percent comprehension were delegation-style, asking the model to solve rather than to explain. Patterns at or above 65 percent involved generating and then interrogating, requesting explanation alongside code, or asking conceptual questions and writing the code independently.

The mature analogue is the aviation and human-factors literature on automation complacency, where the foundational synthesis establishes that complacency and automation bias are the same underlying phenomenon with attention allocation at its center, rather than two separate failure modes.²¹⁰ The transfer to agent supervision is direct: a reviewer whose attention is allocated away from a task by reliable automation does not become a worse reviewer when the automation fails—they become a slower one to notice.

The practical implications are unglamorous. Vary what humans verify so attention does not settle. Seed the review stream with known defects periodically and measure catch rate—this is the approval-integrity probe in Section 20. And require an explanatory record for non-trivial agent-authored change at the time of authorship rather than at the time of confusion, because comprehension debt is invisible until someone needs to change the code.

9.6 Team Topology #

The most credible org-design position holds that organizational maturity rather than model capability determines outcomes, reframing team structure as “infrastructure for agency” with four claims worth carrying:

• bounded agency, where authority to act is intentionally constrained by guardrails so delegated initiative remains governable;
• a security asymmetry question about why organizations grant agents data access they would never permit a human;
• cognitive load transferring to agents, so that agents “begin to lose coherence or hallucinate when they operate outside their defined boundaries” exactly as humans do when their capacity is exceeded; and
• stewardship rather than ownership as the framing for long-lived systems.²¹²

That last set is more than analogy. If agent coherence degrades outside bounded scope—and Sections 5.8 and 7.4 say it does, by measured amounts that worsen with scope—then the Team Topologies cognitive-load principle applies to agents as first-class team members, and scope discipline is a correctness control rather than only a security one.

The same source points to an innovation-and-practices enabling team pattern, and observes that some organizations use “friendly FOMO” rather than mandates to drive adoption. Section 9.7 explains why that distinction matters more than it sounds.

9.7 What Happened at Companies That Mandated It #

Four public arcs, and the pattern across them is instructive.

9.8 Controls #

9.9 Evidence to Request #

• The absorption ratio, trended, with the source of its review-hour estimate.
• Results of the most recent seeded-defect probe, with catch rate and median review time.
• Tenure and attrition data for senior engineers, read alongside review load—the two together are the early warning for depleting the capability everything else depends on.
• Entry-level hiring and internal promotion rates over eight quarters.
• A named owner for a randomly chosen subsystem, and a five-minute explanation of how it works.

9.10 Failure Modes #

• Measuring engineers on AI usage. It is the one policy in this literature with a documented public reversal, and it inverts the relationship between tool and outcome.
• Treating “percentage of code written by AI” as a goal. Undefined, unverified, and an input metric masquerading as an outcome.
• Cutting the junior pipeline while depending on senior judgment. The cohort with the largest measured gains is the cohort being cut, and there is no published mechanism for producing seniors without it.
• Uniform review as policy and waived review in practice. Both failure modes appear in the same dataset; the merge record looks identical either way.
• Assuming a supervision ratio. Nobody has one. Measure absorption instead.
• Believing the agent’s account of the incident. It is the single documented instance where an agent’s self-report was materially false during a live incident, and it is enough to warrant a runbook line.