Agentic Skill Decay

AddyOsmani.com · rss · 2026-08-31 source ↗ Reader 🖥️

Agent Memory Capacity to Verify AI Output Desirable Difficulty & Productive Struggle Domain Expertise as AI Force-Multiplier Learning Culture & Slack Meta-Skills & Contextual Judgment Skill–Ability Gap Systemic Apprenticeship & Skill Transmission Verifiability (as the Quality Bar)

Agentic coding collapses the loop from problem to working patch so fast that it strips out the incidental struggle (debugging, reading diffs, forming and testing hypotheses) that used to build engineering expertise as a side effect of the work itself — so practitioners must now deliberately reinsert that practice, because the very expertise agents make optional is what increasingly determines whether you can verify and direct what an agent produces.

practical · argumentative — Argues for a specific course of action (deliberately reinsert practice into agentic workflows) built on the author's own practitioner experience plus two cited studies, rather than primarily building theoretical understanding for its own sake.

Structure

Part	Role
Reps used to arrive as a side effect of writing code; agents remove that side effect	context / diagnosis
Two abilities agentic work depends on: deep expertise and applied judgment, broken into four practicable skills (deciding, specifying, steering, verifying)	framework
Personal history of building reps the old way (JavaScript, C++, performance optimization) as illustration of how the old mechanism worked	evidence / narrative
The short-circuit: juniors especially can skip the struggle that used to build judgment	elaboration of the diagnosis
A completed task is not necessarily a rep — you must actively extract the lesson	refinement / distinction
2026 Anthropic Trio-library study (50% vs 67% quiz retention) and the ~400,000-session Claude Code study (intermediate expertise raises verified-success odds)	evidence
Four places to slow down: hypothesis before prompting, ask why, read the diff, occasionally work by hand	prescription
Verification is the floor, imagination is the ceiling (illustrated by the 3D album-site anecdote)	framework / evidence
The return on expertise is going up; capture lessons durably (lessons.md, lint rules, tests) so both you and the agent improve	conclusion / prescription

Key terms

reps
The incidental practice — trying an approach, watching it fail, tracing the bug, reading someone else's fix, carrying the lesson forward — that used to arrive automatically as a side effect of manually writing code, and that must now be deliberately re-created.
a completed task
An agent-finished piece of work that, by itself, carries no guarantee of learning — completion and learning are decoupled unless the human actively extracts a lesson from it.
verification (as the floor) / imagination (as the ceiling)
Verification is what bounds what you can safely keep — only what you can confirm is good; imagination is what bounds what you can even ask for — you can only prompt what you can picture. Both are gated by domain expertise.
the outer loop
The residual human job once agents handle execution: writing and refining plans, defining what 'done' means, and deciding where humans must stay in the loop for correctness, safety, or user impact.

Propositions

explicitBefore agentic coding, engineers built expertise as an incidental side effect of writing code — trying approaches, debugging failures, reviewing others' code, reading documentation.conf: high
explicitGood agent-driven work depends on two abilities: deep domain expertise (understanding the problem well enough to define a good outcome) and applied judgment (turning taste into a clear, testable plan).conf: high
explicitAgents let developers skip the struggle (debugging, searching documentation, forming and testing hypotheses) that previously built expertise, so heavy unreflective agent use erodes a person's own skill over time.conf: high
supportsLLMs Reward Expertise · p2 — Both describe the same effect from different angles — agents producing passable output that masks the underlying skill gap of the person directing them.
extendsDet skal være vanskelig · p6 — Independently arrives at Bjork's desirable-difficulty logic (effort is the learning mechanism, not a flaw) and applies it to professional coding practice rather than school assessment.
explicitA completed agentic task does not, by itself, constitute a learning "rep" — learning requires deliberately extracting the lesson (asking why, predicting what could still fail, occasionally working a piece by hand).conf: high
explicitIn a 2026 Anthropic study, junior engineers learning the Python library Trio with AI assistance scored 50% on a follow-up quiz versus 67% for those who worked by hand; within the AI-assisted group, asking conceptual questions and requesting explanations correlated with better retention.conf: high
supportsResearch: Why Some Junior Employees Work Well with AI—and Others Don’t · p8 — Both point to the same mechanism — the entry-level on-ramp that builds judgment through repetition is being displaced by AI assistance — evidenced here by a controlled quiz-retention study rather than a field study of task performance.
explicitVerification is the effective floor of what a person can responsibly ship with an agent (you can only keep what you can confirm is good), and imagination is the ceiling (you can only prompt what you can picture) — both are bounded by domain expertise.conf: high
extendsLLMs Reward Expertise · p9 — Sharpens the general claim that expertise remains structurally valuable into a specific mechanism: expertise functions as the floor (verification capacity) and ceiling (imaginative specification) that together bound what a person can safely ship with an agent.
explicitAn Anthropic study of roughly 400,000 Claude Code sessions found that even intermediate, task-specific expertise increased the odds of reaching verified success, independent of general cross-stack experience.conf: high
explicitLessons learned while working with an agent should be captured in durable, codified forms (lessons.md, lint rules, type constraints, tests) rather than left in chat memory, because chat context and compaction will lose them.conf: high
extendsThe Intent Debt · p22 — Concretizes the same author's earlier 'write down the why' prescription into specific durable-capture mechanisms (lessons.md, lint rules, type constraints, tests) for lessons learned while working with an agent.
explicitThe return on human expertise is increasing, not decreasing, as AI raises the floor of what can be produced, because expertise is what distinguishes shippable, maintainable, high-quality work from merely plausible-looking output.conf: high
implicitAgent throughput scales faster than a single person's attention, so the practical constraint on using agents "aggressively" is the human's capacity to direct and check many parallel sessions, not agent capability itself.conf: medium
implicitUnless deliberate practice is reinserted into the agentic workflow, early-career engineers in particular risk never building the domain expertise and judgment needed to verify AI output — the deskilling risk compounds because the very verification capacity needed to catch it is itself what erodes.conf: medium

Arguments

Practitioners must deliberately reinsert practice — forming a hypothesis before prompting, asking why, reading the diff, and occasionally working a problem by hand — into their agentic workflow. strength: medium

Explicit premises:

Reps used to arrive as a side effect of writing code.
Agents remove that side effect by closing the gap from problem to patch too fast for the old incidental learning to occur.
The four named interventions recreate the moments where learning used to happen.

Implicit premises:

The kind of judgment built through struggle remains necessary going forward and is not being replaced by some other, non-practice-based way of acquiring it.
The practitioner has enough slack/time to spend on deliberate practice, despite the shipping-velocity pressure the author elsewhere names as a real obstacle.

Inferences:

This deliberate-practice burden falls hardest on juniors, who have both the least existing expertise to fall back on and the least organizational standing to ask for the time the prescription requires — the author gestures at this ("going to require junior engineers especially to be proactive") without following it through.

Human expertise is becoming more, not less, valuable as agentic coding spreads. strength: medium

Explicit premises:

Expertise sets the floor (verification) and enables the ceiling (imagination) of what can be safely shipped with an agent.
The ~400,000-session Claude Code study found intermediate task-specific expertise raises the odds of verified success.
Domain expertise is what lets a person notice and diagnose the subtly-wrong output agents produce (illustrated by the 3D album-site anecdote).

Implicit premises:

Verification demand does not itself get automated away by better models or evals at a rate that outpaces the skill-decay problem — i.e., verification remains a durably human bottleneck.

Narratives

The author's own path to expertise — learning JavaScript, C++, and desktop/graphics performance tuning the hard way, including thousands of hours in the Chrome DevTools performance panel before good documentation existed on the topic.
→ Illustrates how reps used to be built through scarcity-driven struggle, and sets up the contrast with agent-run profiling now short-circuiting that same struggle — supports propositions 1 and 3.
Building an interactive 3D album site (3D CD player, vinyl player, tape player): the agent's first pass rendered but used the wrong interaction pattern and performed badly on mobile; the author's own expertise let him notice and diagnose why.
→ Illustrates "verification is the floor, imagination is the ceiling" concretely — expertise as the thing that catches plausible-but-wrong agent output — supports proposition 6.
Running five or ten parallel agent sessions and catching himself having asked the wrong project to add dark mode.
→ Illustrates that agent throughput scales faster than personal attention, motivating the implicit proposition (9) about attention as the real bottleneck to aggressive agent use.

Questions addressed

✅ Does the next generation need to build the same code-level expertise if agents can increasingly do these tasks?
Partially: yes, for now, because agents still don't do a perfect job and someone must notice and diagnose what they get subtly wrong — but the author explicitly declines to say for how long this remains true as models improve.
✅ How do you build expertise once the normal path (reps as a side effect of writing code) is gone?
Four deliberate interventions — hypothesize before prompting, ask why repeatedly, read the diff and predict what could still fail, and occasionally work a problem by hand — plus proactively capturing lessons in durable, codified form.
⭕ What happens to practitioners who lack the organizational slack or permission to slow down and do deliberate reps, given the "velocity expectations" and shipping pressure the author himself names as real?

Gaps

incompleteThe piece never addresses what happens once verification itself becomes more automatable — e.g. agents that critique or test other agents' output at a level exceeding a given human's own judgment. Is the "return on expertise" durable, or does the same skill-decay dynamic eventually reach verification skill itself?
Read next: Work examining whether AI-based evaluation/verification is itself subject to the same expertise-erosion dynamic the article describes for coding, or whether verification is being treated as a stable, un-automatable human function that may not stay stable.
incompleteThe essay is entirely first-person and anecdotal (one senior practitioner's account, backed by two cited studies), and never addresses whether "deliberate reps" can be operationalized for an entire junior cohort at an organizational level, versus remaining an individual discipline mainly available to practitioners who already have the author's level of self-directed rigor.
Read next: Organizational or pedagogical research on structuring deliberate practice for junior engineers at scale, rather than individual habit-formation advice.
illogicalThe author asserts "the return on expertise is going up" with near-certainty, then two paragraphs later states "I don't know how long code-level expertise will remain as valuable as it is today. Models are improving too quickly for much certainty." These two claims sit in tension without being reconciled — the confident framing needs a time-horizon qualifier the piece never supplies.

Closes gaps elsewhere

Det skal være vanskelig · g2 — That gap asks for 'desirable difficulty' to be operationalized — distinguishing effort that builds the mind from mere friction — outside the school-assessment context it was raised in. This article names four specific interventions (hypothesize before prompting, ask why, read the diff and predict what could still fail, occasionally work by hand) as the professional-coding analogue, giving Bjork's principle a concrete, practitioner-facing form.