Why Coding Agents Do Not Eliminate Expertise
A review of Anthropic's research: delegation, verification, and recovery when using Claude Code
Slide contents
1. Why Coding Agents Do Not Eliminate Expertise
A review of Anthropic's research: delegation, verification, and recovery when using Claude Code
2. Product telemetry is not a market law
Original redraw of Anthropic's research frame
3. A large sample still has boundaries
Redrawn from Appendix · Sample · p. 2
4. Claims pass through models and telemetry
Original redraw of the measurement pipeline
5. Nine modes label the dominant work
Redrawn from Figure 1 and Appendix · Work Mode
6. Most sessions are still about code
Redrawn from Figure 1
7. People choose what; Claude chooses how
Redrawn from Figure 2
8. One prompt triggers roughly ten actions
Redrawn from Who decides what
9. Expertise is task-specific, not occupational
Redrawn from Level of expertise
10. Three observable signals stand in for expertise
Redrawn from Table 1 and Appendix · User expertise
11. Expertise and success are measured nearby
Original diagram of possible construct overlap
12. Expert sessions contain over twice the actions
Redrawn from Figure 3
13. Session mix shifts from fixing toward operating
Redrawn from Figure 4
14. Task value is a relative estimate
Redrawn from Figure 4 and Appendix · Task value
15. Success has three evidence thresholds here
Redrawn from Table 2 and Appendix · Derived measures
16. The largest gap separates novices from competence
Redrawn from Figure 5
17. Expertise shows more clearly under trouble
Redrawn from Figure 5 · troubled sessions
18. Success tracks task knowledge more than occupation
Redrawn from Figure 6
19. Validation supports the gradient, not causality
Redrawn from Appendix · Validation and robustness
20. The study shows association, not cause
Supported
Expertise tracks outcomes
Experts delegate more
Gradient survives controls
Not established
Training causes the same effect
More actions mean better work
Result generalizes to all agents
21. Expertise may expand delegation bandwidth
Original model: framing × verification × recovery
22. Enablement should teach judgment, not syntax
Build in people
Precise task framing
Knowledge of invariants
Diagnosis and recovery
Provide in platform
Domain context and vocabulary
Checks by default
Observable artifacts
23. Evaluate outcomes, delegation, and recovery
Original scorecard derived from the report's limits
24. A pilot should separate input from outcome
Measure expertise before the task
Review artifacts blind
Label trouble from telemetry
Check outcomes after adoption
25. Leaders need a new work contract
People own
Goal and constraints
Meaningful verification
Decisions under trouble
Platform owns
Accessible context
Provable results
Boundaries and rollback
26. Track whether the expertise premium shrinks
Verified-success gap narrows
Novices recover more often
Planning shifts to the agent
Quality remains stable
27. Agents absorb execution, not judgment
People still decide what
Competence captures the main gain
Expertise supports recovery
Measure delegation bandwidth together with quality.
28. Expertise becomes delegation bandwidth
Not more activity. More verified work per unit of human judgment.
Book Cube
The primary report and appendix are linked on the first slide. Continue the discussion in Book Cube.
Alexander Polomodov, Technical Director & Fellow, T-Technologies
@Book_Cube