Coding Agents

By 2025-2026, coding agents — LLMs with file-system access, terminal-execution, and project-wide code understanding — are the most-deployed agentic-AI application. The lineage runs from Codex / Copilot (inline completion, 2021) through chat-based code Q&A (2023) to autonomous edit-test-fix loops (2024-2025). Cursor, Claude Code, Codex CLI, Aider, GitHub Copilot Workspace, Devin, Continue, Cline, Windsurf, Roo Code — the market is large, fragmented, and rapidly improving.

What a coding agent does

The standard capabilities of a 2025-2026 coding agent:

• Read project files — full filesystem access scoped to the project root.
• Edit files — surgical edits, multi-file refactors, new file creation.
• Run terminal commands — build, test, lint, format, install dependencies.
• Execute generated code — run scripts, REPLs, test suites.
• Search across the codebase — grep / ripgrep / semantic search.
• Git integration — read history, view diffs, sometimes commit and push.
• MCP / external tools — increasingly, coding agents call out to external services (databases, web search, deployment platforms).

The interaction loop is essentially ReAct: plan, take an action, observe the result, plan the next action. Modern coding agents iterate hundreds of times per session.

The product landscape

Inline completion: GitHub Copilot, Cursor Tab, Codeium

Direct descendants of Codex/Copilot. Ghost-text completion in the IDE, low latency, no chat. The default for high-volume per-keystroke assistance. By 2025 these have absorbed structured-edit features (multi-line refactors, "Tab to apply").

IDE-integrated agents: Cursor, Windsurf, Continue, Zed

Cursor (founded 2022) defined the 2024-2025 "AI-native IDE" category. Tab completion + chat + agent-mode that can edit multiple files. Cursor's "Composer" and "Agent Mode" are reference implementations of in-IDE agentic editing. By 2025, Windsurf, Zed, Continue, and JetBrains AI compete in the same space.

CLI agents: Claude Code, Codex CLI, Aider

A different product shape: command-line agents that operate on a project directory.

• Aider (paul-gauthier, late 2023) — open-source pioneer. Edit-mode LLM with git integration.
• Claude Code (Anthropic, Feb 2025) — official CLI agent. Tightly integrated with Claude's tool-use; runs in any terminal.
• Codex CLI (OpenAI, April 2025) — OpenAI's answer to Claude Code.

CLI agents tend to be preferred by experienced developers who want full control over the environment. They also dominate for long-running tasks (hours of edits) where the IDE-chat pattern is awkward.

Autonomous agents: Devin, OpenAI Agents, Gemini Coder

The "give it a goal, walk away" category. Devin (Cognition, March 2024) was the first to ship publicly under this framing. By 2025 most major providers offer some variant. Capabilities at the high end: take a Linear / Jira ticket, propose a plan, execute with the codebase, open a PR, respond to review comments.

Reliability remains the key gating factor — long-horizon autonomous coding still fails on enough tasks that human supervision is required.

Benchmarks

The benchmark landscape for coding agents:

• HumanEval, MBPP — saturated since 2024.
• SWE-bench Verified — real GitHub issues with tests. Frontier models scoring 60-75% in mid-2025.
• LiveCodeBench — competitive-programming problems updated continuously to avoid contamination.
• Aider's polyglot benchmark — multi-language editing.
• TerminalBench / OSWorld-Code — agentic terminal-use evaluation.

SWE-bench Verified is the most-watched coding benchmark in 2025-2026. Frontier-model performance on it is a primary signal about coding-agent capability.

What's actually working

Tasks coding agents reliably do well:

• Boilerplate generation — new files, typed-stub APIs, configuration scaffolding.
• Refactors with clear specifications — rename, extract function, change a type signature consistently.
• Bug fixes given a stack trace — particularly when the cause is local.
• Adding tests to existing code.
• Code review — pointing out issues, suggesting improvements.
• Documentation generation.

Tasks that remain hard:

• Architectural decisions — "should this be a microservice?" The agent will pick a design but reasoning about long-term implications is weak.
• Subtle bugs crossing module boundaries.
• Performance optimisation requiring profiling and runtime knowledge the agent lacks.
• Reading large codebases for high-level understanding without prior familiarisation.

How developers are using coding agents

Survey data from 2024-2025:

• >70% of professional developers use AI assistants daily as of 2025 (StackOverflow, GitHub surveys).
• Inline completion is the most-used feature; chat-based and agent-based features grow rapidly.
• Productivity gains are real but contested in size — studies show 10-50% speedup on certain tasks, no gain (or regression) on others.
• Skill effects are debated — junior developers may benefit more, but the "skill-erosion" concern (developers losing ability to code without AI) is increasingly raised.

What this changes

• Codex / Copilot's inline completion is now table-stakes; the differentiation has moved up the stack to agentic capability.
• Open-source coding agents are competitive with closed-source for many tasks (Aider + open-source models from Qwen, DeepSeek, Mistral).
• Coding-agent capability has pulled ahead of general LLM capability on real-world software tasks — providers tune specifically for coding workflows.
• The skill premium for senior developers has shifted toward those who can effectively direct AI agents, not write all the code by hand.

What to read next

• Codex & Copilot — the lineage starts here.
• Tool Use — the protocol coding agents use.
• Computer Use — agentic action beyond just code.