Coding Agents

Definition

AI systems that autonomously write, execute, test, and iterate on code — going beyond code completion to full agentic loops: read problem → write code → run → observe result → revise.

Current State of the Art (2026)

  • Claude Code (Opus 4.6): the leading general coding agent as of early 2026
  • Codex (OpenAI): used by OpenAI internally via harness engineering
  • SWE-bench / merge rate as evaluation (see Are LLMs Not Getting Better?)

What Works

Autoresearch and optimization: With access to GPU clusters, coding agents can run 910+ experiments in 8 hours, discovering interaction effects that sequential search misses (Scaling Autoresearch). They discover emergent strategies (heterogeneous hardware routing) without explicit instruction.

Algorithm optimization: Claude Code matched published SOTA on algorithmic problems without evolutionary scaffolding, given only problem + solution + evaluator (AutoEvolver).

Architecture analysis: A coding agent reverse-engineered Claude Code’s architecture in one day using multi-agent orchestration (Claude Code Analysis).

What Doesn’t Work (Yet)

Merge rate stagnation: Statistical analysis shows LLM merge rates (code maintainers would approve) haven’t improved since early 2025 (Are LLMs Not Getting Better?). Test passage ≠ real quality.

Cognitive debt: Agents generate code faster than humans can understand it. “Tiny booboos compound at a rate that’s unsustainable” (Slowing Down). Speed removed the human cognitive bottleneck but didn’t replace it.

Aspiration satisficing: Agents declare victory at local optima. Need explicit “aspiration prompting” (tell the agent the SOTA is X) to break through (AutoEvolver).

Design Patterns

  • Harness engineering: Encode quality discipline into the environment (tests, linters, automated checks) rather than relying on agent judgment (Harness Engineering)
  • Parallel factorial search: More GPUs → factorial grids → catch parameter interaction effects (Scaling Autoresearch)
  • Tool categorization by side effects: Read-only tools parallel, write operations serialized (Claude Code architecture)
  • Agent Interaction Guidelines: Transparency, disclose agent identity, keep humans accountable (Linear AIG)

Open Questions

  • Is the merge rate plateau a ceiling, or did it step up with Claude 4 / Gemini 2.5?
  • What’s the right human-in-the-loop cadence for different task types?
  • Can aspiration prompting scale to complex multi-month engineering projects?