Anthropic Claude Code engineer Addy Osmani argues that heavy Coding Agent users already manage work—picking tasks, writing briefs, tracking progress, and deciding what ships. The edge is not prompt craft alone, but making agents self-verify while keeping product and release judgment with humans.
Background: cheaper code, more orchestration
On November 8, Addy posted *Your Coding Agents Made You a Manager*. A day can start by checking several agents: done work, failed tests, PRs awaiting review, tasks to reassign. AI writes much of the code, yet human load shifts from typing to orchestration and acceptance—close to the IC path many engineers tried to avoid.
That echoes this site’s talk with Claude Code desktop lead Felix: set goals that feel unreasonable, then make the model treat them as real. Goals and verification beat piling on prompts.
Make the agent prove the work is done
Strong teams use CI, tests, and review—not managers reading every line. Same playbook for agents:
– Self-verification first: ship test suites, builds, linters; for UI, let the agent open a browser. Put commands in CLAUDE.md; use `/goal`, Stop Hooks, and a separate `/code-review` sub-agent so “looks done” is not “done.” – Add only what the repo lacks: why it matters, constraints, success criteria—skip structure the model can discover. – Compound corrections: after the same fix twice, write it into CLAUDE.md or a Skill; prune stale rules so the instruction file does not become new debt. – Less micromanagement as capability rises: enforce non-negotiables with Hooks / Deny Rules; leave preferences to the agent. Endless “think harder” rules often recreate human micromanagement.
Delegate tasks, not judgment
For hard work, start in Plan Mode; optionally have a second session critique the plan. `/batch` can split large changes into 5–30 parallel tasks, but humans must police boundaries to avoid interface collisions. Correct early; after two failed corrections, clear the session and rewrite the prompt instead of stuffing a bad context.
Addy’s cautionary case: an agent researched competitor gaps and opened a ~600-line, all-green PR he nearly merged—then realized limited user value and extra complexity. Passing tests ≠ worth shipping. Style and routine fixes can go to agents; product tradeoffs, security/compliance, and go-live stay human. Keep review units small enough for a person to actually understand.
Multi-agent runs also incur an “orchestration tax”: Addy often runs four or five background agents before attention becomes the bottleneck. Scale automatic checks and alerts, not babysitting. Measure cycle time, rework, and rollbacks—not just lines or PR count—or “productivity” merely moves from coding to the review queue. The same caution shows up in our note on SwarmEvolve’s multi-agent evolutionary search: scale alone is not enough; collaboration and acceptance decide effective output.
Managing agents ≠ managing people
Agents make cheap A/B implementations possible, but cross-session memory depends on saved notes and shared team knowledge—not a substitute for long-horizon human judgment. Models also tend to comply; with “cognitive surrender” risk, ask the agent to challenge your conclusions. Junior practice opportunities and human trust still matter when agents absorb entry-level work.
Addy’s point is not “everyone become an EM.” Staff engineers already set constraints, design, and accept work; Coding Agents simply push those skills to more people. To get full value, learn to manage the agents—without changing your job title.
FAQ
Q: Do more Coding Agents mean linear speedups? A: Not necessarily. Parallelism saves machine time; human coordination and review create an orchestration tax that can bottleneck on the PR queue.
Q: Are all-green tests enough to merge? A: No. Automation checks correctness; whether the change is worth shipping, plus security and compliance, still need human judgment.
Q: Should CLAUDE.md keep growing forever? A: No. Capture reusable verification commands and recurring pitfalls, then prune stale rules so micromanagement does not slow the agent.
相关文章
