Context Is the Scarce Resource

PM craft when your executors — human or agent — go fluid

The most interesting management shift happening right now is invisible if you only look at it from one direction.

Look at it from the AI side and you see PMs starting to direct AI agents the way they once directed teams: writing system prompts that read like PRDs, orchestrating handoffs between specialized models, sampling outputs for quality the way they used to sample sprint deliverables. The medium is new; the craft is recognizable.

Look at it from the human side and you see something else. In her recent piece on the topic, Cassie Kozyrkov calls it chimeric talent: people who fuse their core expertise with on-demand AI extension, picking up SQL in a morning and shipping production code by next week, building breadth in hindsight rather than over years. T-shaped professionals took years to add capability around their depth. Chimeric workers, in her framing, reach for whatever the task needs and synthesize it on the spot. Time-to-competence collapses from months to hours. Job architectures calibrated for stable expertise turn brittle.

These two shifts are usually discussed as separate phenomena. They aren't. They're the same shift hitting management from opposite directions, and the synthesis reveals what the work actually is now.

The real distinction isn't human vs. agent

Once executors — human or AI — can mutate capability fluidly, the population you're managing stops sorting cleanly by substrate. A senior IC with eight years of context and a model with a great system prompt and the right retrieval are both high-context, capable executors. A new hire reflexively pasting tickets into Claude without reading the codebase and a freshly-spawned agent with no system prompt are both low-context, fast, and fluent enough to mask it.

The variable that matters isn't whether the executor is silicon or carbon. It's how much situational knowledge they're operating with — and whether the work they're doing is calibrated to that level.

The dangerous failure mode here is the high-skill worker running at lightning speed straight into the organizational wall, automating the wrong thing, optimizing for metrics that don't matter. Kozyrkov's diagnosis is sharp: skilled but context-blind work is the new risk, not lack of capability. That's the same failure pattern as an agent producing a beautifully formatted answer to a question the user didn't actually ask. In both cases the executor doesn't know what it doesn't know. In both cases the failure is invisible without a reviewer carrying the context.

The bookend failure mode — what Kozyrkov names workslop — is skill applied without care, polished-looking garbage shipped because the executor didn't bother to verify. Same dynamic, polarity reversed.

Either way, the management problem isn't humans versus agents. It's a portfolio of executors varying along two axes — capability and context — and the job is routing work to the right cell of that grid while adjusting autonomy as context accrues.

Context is the scarce resource

Kozyrkov puts the central reframe cleanly: "Skills are abundant now. Context is scarce." AI gives anyone capability — write SQL, generate a design, scaffold an API. What it doesn't give them is the knowledge that customer X drives 40% of margin, or that you tried this exact GTM in 2023 and it failed because of a partnership constraint that lives in nobody's deck.

The same is true at the agent level. A model can write the code; what it can't do without help is know which branch of the codebase is canonical, why a particular abstraction exists, or what the team's unwritten taste is. The "context window" isn't really a token-count problem — it's a question of whether the executor has the situational knowledge to make a non-obvious call correctly.

Context curation becomes the highest-leverage thing a PM does. What's the canonical source of truth? What's the unwritten rule that explains why you don't do X? What does a new joiner — human or agent — need in week one to be productive without being dangerous? Writing a great onboarding doc, a great PRD, and a great system prompt is fundamentally the same discipline. The leverage just got more obvious.

Trust calibration is the central skill

If context is what makes capability productive, what Kozyrkov calls trust calibration is what makes management work. The skill is figuring out — faster than ever — who can be trusted with what level of autonomy. Sample work strategically. Check high-stakes decisions closely. Let routine work from proven operators flow. Watch how people handle correction. Test judgment, not just output. Her shorthand for the new manager's instinct: ask how someone approached the problem, not just what they produced.

This is the same loop a good agent operator runs. You don't review every line a model generates; you sample by risk. You design the surface area of autonomy — what tools it can call, what decisions require approval — based on demonstrated reliability. You watch for the specific failure mode of confident, fluent output that solves the wrong problem.

The discipline ports across substrates because the underlying question is the same: given an executor who can produce more than I can review, what's the smallest, smartest sample that tells me whether to trust the rest?

Managers who can't articulate what they're sampling for will either over-review and kill throughput, or under-review and eat the workslop. Neither scales.

The 10x problem, squared

The classic 10x-engineer problem — a single hyperproductive worker destabilizing a team of 1x peers — gets multiplied when AI gives the 10x-er another order of magnitude of throughput. Kozyrkov's version is that AI 10xes the 10x problem itself. Now one chimeric IC outproduces five colleagues, breaking career ladders and pay bands, asking a model rather than a senior coworker when they hit a wall. The same dynamic plays out when one PM is directing ten agents instead of two.

The destabilization isn't really about output volume; it's that the team's interfaces, review processes, and trust contracts were calibrated for the old throughput. Once one node is running ten times faster, every interface it touches becomes a bottleneck candidate — review capacity, ambiguity resolution, stakeholder communication. PMs operating in this world spend more time redesigning the team's review surfaces than they do directing any individual unit's work. The fix is to redesign the interfaces around the new throughput, not slow the fast unit down to match what the system can absorb.

What follows

The PM role tilts further toward context engineering: writing canonical docs, designing onboarding for both humans and agents, making unwritten rules legible, curating what enters the working set of any given decision-maker. This was always part of the job. It becomes the highest-leverage part.

Trust calibration becomes a teachable skill rather than a vibe. The instinct to ask how someone approached a problem, not just what they produced, applies to direct reports and agent runs alike. Both require articulating the failure modes you're sampling for.

Hiring shifts toward judgment and curiosity over credentialed skill. If skills can be borrowed in an afternoon, the durable signal is whether someone asks the right questions before acting, updates their model when corrected, and recognizes what they don't know. The same heuristics apply to evaluating which agent setup to adopt.

Context preservation becomes infrastructure work. If senior humans hold most of an organization's context and juniors increasingly converse with models more than with peers, institutional memory has nowhere to live by default. Making context queryable, transferable, alive — that's not a nice-to-have anymore; it's the substrate that determines whether any of the new throughput is actually productive.

The unifying thread: AI doesn't reduce the need for management, it relocates it. The questions move from do you have the skill? to do you have the context? — and from how do I direct this work? to how quickly can I figure out who, or what, to trust with how much?

Capability is now cheap. Context is the moat. The work is learning to tell the difference.