Source and takeaway
Eleven production coding agents, about 4M lines between them, all write their own loop: none imports a general agent framework and none uses embeddings to search code. That is the core finding of Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents (Barbaste, Darrigol, Vu and Wiltberger; arXiv 2609.00006, cs.SE, 83 pages), the expanded second edition of an April 2026 study.
It reached us through a LinkedIn post by Dinesh Reddy (about 2026-11-10; 59 reactions, 33 comments), which compressed it to one line: the framework isn't the product, the harness is. The post cites the paper by arXiv number only and links nothing.
For a staff engineer the useful part is not the slogan. It is the paper's map of seven harness subsystems, the smallest and largest version of each it found, and 18 concrete recommendations. Read it as a design survey from source code, not as a benchmark: it measures nothing at runtime.
What the study did
The authors read the source of eleven coding harnesses and dissected each along the same seven subsystems. The result is 13 cross-cutting observations, 29 recurring design patterns and 18 recommendations. The corpus grew from eight systems in the April edition, and the retained April snapshots give a 90-day before-and-after diff of the same codebases.
The eleven: Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode and OpenClaw. Omnigent, a meta-harness that coordinates other harnesses, is a contrast case outside the eleven.
| Subsystem | What it covers | Notable range in the corpus |
|---|---|---|
| Loop | The turn cycle: call the model, run tools, decide to stop | Mini-SWE-Agent's ~100-line loop up to multi-mode loops with plan mode, hooks and recovery |
| LLM integration | Provider clients, streaming, retries, model routing | Single provider up to many providers behind one interface |
| Tools | Built-in tools, how they load, how results return | A bash tool alone up to on-demand tool loading past ~15 tools |
| Memory and context | What enters the window and how it is found | ripgrep, glob, tree-sitter and Markdown context files everywhere; embeddings only for chat memory (OpenClaw) |
| Safety | Approval modes, sandboxing, policy | Full OS sandbox plus policy-as-code down to none (Hermes and OpenCode are large but unsandboxed) |
| Orchestration | Sub-agents, parallelism, hosting other harnesses | Sub-agents mostly in-process; ACP lets one harness host another |
| Extensibility | Skills, MCP, hooks, plugins, SDKs | Skills in 9/11, MCP in 8/11, ACP in 6; Pi rejects MCP |
Key findings
No frameworks, no code embeddings. Across ~4M lines, no harness imports a general agent framework; Gemini CLI skips Google's own. Code is found with ripgrep, glob, tree-sitter and Markdown context files.
Loop size predicts production readiness, not benchmark score. Mini-SWE-Agent (~100 lines) scores on SWE-bench in the same range as systems a thousand times larger. The extra code buys safety, recovery and extensibility.
Skills lead MCP. 9 of 11 ship skills, 8 of 11 MCP. Pi rejects MCP outright; OpenCode reads
~/.claude/skills, so the skill format is becoming portable.ACP has a new job: harness hosting. OpenHands can run Claude Code, Codex or Gemini CLI as backends. Sub-agents still mostly run in-process.
Safety is a choice, not a function of size. Two large systems (Hermes, OpenCode) have no OS-level sandbox.
Convergence has become imitation. In 90 days Codex adopted Claude Code's hook vocabulary and plan mode, features spread within weeks, and behavioural rules moved from prompt text into configuration. Codex roughly doubled to ~1.1M lines of Rust.
The harness is now the platform. Harnesses ship as SDKs (Claude Agent SDK, openai-codex, openhands-sdk). Those SDKs are, in effect, the new framework layer, which complicates the post's "no frameworks" reading.
The recommendations, condensed
The paper lists 18; these are the ones with a concrete design consequence.
Start with a while loop and a bash tool; add subsystems only when a failure demands them.
Load tools on demand once you pass about 15.
No retrieval-augmented search over code unless it beats ripgrep plus tree-sitter on a held-out eval.
Give developer tools three approval modes: PLAN, DEFAULT, YOLO.
For enterprise use, an OS sandbox plus policy-as-code.
Stay single-agent until parallelism clearly pays.
Ship an ACP server so other harnesses can host yours.
Skills before MCP.
No LangChain-style runtimes.
Cheap caps on stuck loops.
The paper closes with a 90-line minimal harness in Python as the starting scaffold.
Tradeoffs and failure modes
The grep-first loop is the right default, but it has known limits, and the post's own comment thread names them.
Text edits break semantic refactors. A grep-driven rename in Java hits comments and strings and misses overloads. Where the language has a compiler-accurate refactoring tool, the edit should go through it.
Agents drift back to grep. Commenters report that agents given LSP or compiler-backed tools over MCP still fall back to ripgrep. Offering a better tool is not enough; the prompt, tool descriptions or a gate has to steer to it.
grep needs a map. In a large repo nobody has documented, lexical search finds strings, not structure. The Markdown context files every harness uses are what make grep work; without them the loop wanders.
Mid-run tool failure. Several commenters ask how a minimal loop recovers when a tool call fails halfway. This is exactly what the larger harnesses spend their code on: retries, checkpoints, stuck-loop caps.
"No framework" is partly a naming question. Each harness is itself a framework, and now ships as an SDK. The real claim is narrower: generic orchestration runtimes added cost without adding capability for coding agents.
Caveats
Static reading only. The study reads source code; nothing is measured at runtime, and there is no head-to-head run of the systems.
Qualitative scoring. Subsystem ratings are the authors' judgment.
Weakest on Claude Code. That analysis rests on an unofficial March 2026 source snapshot, which the authors call the least reproducible part.
The SWE-bench point is borrowed. "Same range as systems 1,000x larger" relies on published scores, not the authors' runs.
Moving target. A 90-day diff already shows large shifts, so specific counts (9/11 skills, 8/11 MCP) will age fast.
Metadata oddity. arXiv lists v1 on 15 Jul 2026 under a 2609 identifier; cite it by ID.
Review checklist for a harness
Is there a generic agent framework in the dependency tree, and what does it do that a while loop does not?
Does any code search use embeddings? If so, is there a held-out eval showing it beats ripgrep plus tree-sitter?
Are there Markdown context files (AGENTS.md, CLAUDE.md or similar) that give grep a map of the repo?
Do semantic edits (renames, signature changes) go through a compiler-accurate tool, and does something stop the agent drifting back to text edits?
More than ~15 tools? Are they loaded on demand?
Are approval modes explicit (plan, default, auto) rather than implied by prompts?
Is there an OS-level sandbox, and are policies config rather than prompt text?
Is there a cheap cap that detects and stops a stuck loop?
Is a second agent justified by measured parallel gain over one agent?
Are capabilities packaged as skills first, with MCP only where a live service is needed?
Can another harness host this one (ACP), and can it host others?
Related entries and sources
R006 6 must-know terms for AI agent harnesses: the vocabulary; this study is the evidence behind it.
R023 Building a coding agent from scratch: a buildable harness that starts where the 90-line scaffold does.
R022 Six-stage AI engineering reading list: Anthropic and OpenAI harness-engineering posts; same "re-test each component" stance.
R010 PageIndex: retrieval as tools without embeddings, for documents instead of code.
R044 Formal Skill: what skills could become if the skills-first trend holds.
R030 OpenShell: the OS sandbox plus policy the paper recommends for enterprise use.
R019 s3 agent swarms: the multi-agent case to test against the paper's single-agent default.
R034 AG-UI: another agent protocol; ACP is the harness-to-harness slot.
Sources: arXiv 2609.00006; Dinesh Reddy's LinkedIn post.