What it actually does
Claude Code runs from the terminal. You give it a task, it plans the work, edits the files, runs commands, and verifies. The agent loop is genuinely tight: plan, edit, run, observe, revise. Compared to IDE-native agents like Cursor, the terminal-native shape is what makes it good at multi-file refactors and at long-running tasks where the human is supervising rather than driving.
I have run Claude Code across roughly 60 of our ~500 production sites for the last six months. It is the strongest model-grade coding agent I have used and it is also the most expensive when you let it loose unsupervised. The trick is the structure you put around it before you hit run, not the prompt you write at run-time.
What is good
- Multi-file reasoning.Edits across 5+ files in one task without losing context. Cursor's Composer has caught up here, but Claude Code's terminal-native shape gives it more room.
- Verification discipline. The model genuinely tries to verify its own work: runs the tests, checks the build, retries on failure. We catch unverified output less often with Claude Code than with comparison tools.
- Honest about uncertainty. When the model does not know, it says so. The other coding agents we have used hedge less; that hedge matters in production because hedging is information.
What is broken or surprising
- Cost cliff under loose supervision. Run Claude Code on a vague spec, walk away, and you can rack up a serious bill on a single task. Run the same spec with a tight planning step in front, and the cost is 30% of the loose-supervision version. The pattern is consistent.
- Tool sprawl. The model will reach for tools it does not need if the toolset is wide. Constrain the available tools per task. We have a fixed-toolset launcher script for this.
- Silent context drift on long tasks. On tasks that run more than 20 minutes, the model occasionally produces work that is internally consistent but no longer aligned with the original spec. This is the Silent Drift failure mode from the Failure Pyramid. We mitigate by checkpointing the spec mid-task.
When you would choose it
Pick Claude Code if your work is multi-file, terminal-friendly, and you can put a planning step in front of the model. Pick it if you want the strongest single-model coding agent in 2026 and are willing to instrument it carefully.
Skip Claude Code if your work is in-IDE inline edits and tab-completion (Cursor wins). Skip it if your team will not put structure around it; the cost spike will hurt more than the productivity gain helps. The honest comparison rule lives at claude-code-vs-cursor.
Cost at scale
Across 60 of our sites running Claude Code in our build pipeline, our P50 cost-per-task lands in the low-cents range. P95 is roughly 3x P50. P99 is the cost cliff: 8-10x P50 on tasks where the model spawned excessive tool calls. This is consistent with the orchestrator-worker cost cliff from the patterns essay; Claude Code is, internally, an orchestrator-worker pattern.
The structural fix that flattened P99 in our pipeline: a planning step before any task, where a smaller cheaper model produces a constrained plan that Claude Code then executes. The plan is the cap. P99 dropped roughly 60% after we shipped this. Subscription pricing is fine for individual use; for pipeline use, treat it as a per-task cost line and budget accordingly.
Read next
Oliver runs Digital Signet, a research and product studio that operates ~500 production sites with AI agents as the engineering layer. The Digital Signet portfolio is built using a continuous AI-agent build pipeline, one of the largest agent-operated publishing operations on the open web. The handbook draws directly from those deployments: real cost data, real failure modes, real recovery patterns.