Agentic coding

The agent plans, edits, runs tests and iterates. The human sets goals and reviews.

Also called: Agentic engineering, coding agents

A coding agent takes a scoped task, reads the codebase, makes multi-file changes, runs commands and tests, and iterates until it believes the task is done. The human sets the goal, supplies context and conventions, approves risky actions and reviews the diff before it merges.

Karpathy later preferred "agentic engineering" as the framing for serious work, as distinct from vibe coding.

Agents that could edit code and run tests appeared in research first, including Princeton’s SWE-agent and Cognition’s Devin in 2024. Terminal-based agents such as Claude Code, OpenAI Codex and Gemini CLI followed in 2025 and brought the approach into everyday use.

Karpathy proposed “agentic engineering” in February 2026 and set out the idea at Sequoia’s AI Ascent in April. Others, including Simon Willison and Addy Osmani, had used the phrase earlier.

↺ Steps 2–4 repeat inside one task.

1. 01Write a scoped task with context: the issue, the goal, and a conventions file in the repository (such as CLAUDE.md or AGENTS.md).
2. 02The agent plans, reads the code, edits files and runs commands and tests.
3. 03Permission rules decide which actions need human approval.
4. 04The agent iterates until tests pass, or reports what it couldn’t do.
5. 05A human reviews the diff and merges.

Typical tools: Claude Code, OpenAI Codex, Gemini CLI, Cursor agent mode, GitHub Copilot coding agent.

Compare all five bands

Solid: AI does core work · Hatched: partial or informal · Dashed: done by people. Compare all bands

Good for: Multi-step internal features with a mostly clear goal

Less suited to

• Goals you can’t yet state clearly
• Codebases without tests, where “done” can’t be checked
• Sensitive environments without sandboxing or permission controls

Signs you’ve outgrown it

• Diffs are too large to review properly
• Several people or agents work on the same system
• Requirements drift from one task to the next

Where it runs: A terminal or editor agent (Claude Code, Codex, Cursor) working in a repository with tests.

Isolation

A container, cloud dev environment or separate branch per task

Permissions

Allow, ask and deny rules for commands; no production credentials; limited network access

Audit

Logs of agent actions

On campus

HPC clusters with agents as installed software (Utah, USC); cloud dev environments run by central IT

Compare environments across bands

• A human reviews every diff before merge
• Version control and automated tests
• A separate development environment with scoped credentials

Risks

• Destructive actions: deleting files or running commands on shared systems.
• Prompt injection: instructions hidden in files, issues or web pages the agent reads.
• Credential exposure: agents reading or echoing secrets.
• Review bottleneck: agents produce changes faster than people can review them.

1. Yang et al., “SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering,” 2024
2. Anthropic, “Claude Code: Best practices for agentic coding,” 2025
3. Karpathy on “agentic engineering,” Feb. 2026, and Sequoia AI Ascent talk, Apr. 2026
4. USC CARC, “AI Coding Agents” user guide, 2026