Hold the model constant and change only the software around it, and its score moves by ten points. That is the finding that reorganized the coding-agent market in 2026 — and it is why the real contest now is between harnesses: Claude Code, OpenCode, Meta's Muse Code, DeepSeek Harness, Qwen Code, and Pi.
October 21, 2026 · 8 min read
Start with the measurement, because it disciplines everything after it. Particula Tech ran the same Claude Opus 4.5 through four agent frameworks on SWE-bench Pro: a standardized scaffold scored 45.9%, Cursor 50.2%, Auggie 51.8%, and Claude Code 55.4% — a 9.5-point spread with the weights held perfectly constant. Scale AI's own analysis puts the harness swing at 10 to 20 points. LangChain, changing nothing but its harness, moved a fixed model from 52.8% to 66.5% on Terminal-Bench 2.0. The frontier models at the top of SWE-bench Verified sit within a point of each other; the scaffolds around them do not. Birgitta Böckeler's formula is the one to keep: Agent = Model + Scaffolding — and the scaffolding is now the product. Here are the five harnesses that define the field, and what each one is actually betting on.
Anthropic's entry is really three products in one binary: a terminal agent, a harness specification (tools, permissions, subagents, skills with progressive loading, hooks that intercept events), and an Agent SDK that exposes the same harness to embedders. It is the maximalist design — Agent Teams that share task lists and message each other, an Agent View fleet dashboard, fork subagents on by default since mid-October — and the evidence base is the deepest in the field. The top Terminal-Bench 4.0 pair in October is Claude Code running Opus 5.5 (64.85%, official board, Oct 6), and in Particula's controlled run its scaffold extracted more from Opus 4.5 than any rival. The constraints are the mirror of the strengths: Claude models only, proprietary, and priced through Anthropic's subscriptions or API. If you want the best-measured default and live inside one lab's models, this is it. Mario Zechner's jab — that it is "a spaceship packed with 80% functionality I never use" — is also true, and it is the opening everyone below is exploiting.
OpenCode (Anomaly, formerly SST) is the open-source breakout: MIT-licensed, ~212,000 GitHub stars at v1.19 in mid-October, with version 2.0 shipping October 14 — one background service powering terminal, desktop, and web, background subagents, and a Code Mode that chains MCP tools in a single script. Its bet is provider neutrality as a feature: 75+ providers through Models.dev, local models via Ollama, mid-session model switching with context intact, a read-only plan agent beside the full-access build agent, and LSP integration that loads the right language server automatically. The politics are part of the product: Anthropic forced it to drop Claude Pro/Max subscription login in early 2026 (API keys still work; ChatGPT subscriptions log in natively), which is precisely the lock-in its users are fleeing. Two honest caveats: you are responsible for picking the right model per task, and its hosted side made bad news on October 17 — a breach of OpenCode Console that may have exposed stored provider API keys. The company is clear, and the reporting agrees, that the breach hit the hosted Console, not the open-source agent — but it is a reminder that "open harness" and "hosted convenience" carry different risks.
Meta arrived late and aimed at the cost curve. Muse Code (public beta, August 5, 2026, macOS and Linux) is a terminal agent built on Muse Spark 1.2 — a model co-trained alongside the harness itself, the same move DeepSeek made with post-training and Ornith made with scaffold optimization: stop treating the scaffold as an accessory. Its headline mechanics are parallelism and durability: for large jobs it fans work out to sub-agents running simultaneously in isolated git worktrees so concurrent edits cannot collide (Zuckerberg's launch demo built six game features at once), and every model call, tool invocation, approval, and edit is appended to a local event log that Meta describes as replay-exact and restart-safe — crash mid-task and it resumes rather than restarts. The pricing is the strategic tell: pay-as-you-go at $1.25 per million input tokens, undercutting the US labs the way Chinese labs undercut them on model price. The unknowns are equally clear: a beta, one vendor's model, and benchmark evidence that is so far mostly Meta's own.
DeepSeek looked at a features race and answered with a runtime. DeepSeek Harness (DSH), announced in August 2026, is MIT-licensed, Node.js on the Cordis composition framework, and built on one idea: everything is a plugin — the model adapter, the tool registry, the session log, and the agent loop itself are replaceable components with no privileged core. It ships four modes (Standard, Minimal, Code, Creator) across seven provider families, an append-only event log that gives you resume, fork, and replay, real sandboxing (Landlock on Linux, Seatbelt on macOS, restricted tokens on Windows), and — tellingly — Claude Code and Codex built in as subagent providers, off by default. The Minimal mode is not a demo: it is what runs DeepSeek's official evaluations, and the co-design shows in the numbers — the V4-Flash-0731 release, architecturally identical to its preview, jumped from 61.8 to 82.7 on Terminal-Bench 2.1 through post-training aimed at agentic work. One secondary roundup notes DSH passed OpenCode in GitHub stars within weeks of launch; treat star counts as weather, not climate — but the direction of travel, a first-party open runtime from the lab that prices everyone else into a corner, is hard to miss.
Alibaba's entry has the most instructive lineage in the field: Qwen Code is a fork of Google's Gemini CLI — same architecture, same TUI, same project-context system — with Qwen3-Coder as the default brain and Alibaba's Model Studio (DashScope) behind it. The fork now outlives the original's consumer life: Google switched off Gemini CLI for individual accounts in June 2026 (see round two), while Qwen Code, Apache 2.0 at QwenLM/qwen-code, kept shipping — v0.22.0 stable in August 2026, nightlies into September, roughly 27,600 stars. It is model-agnostic in the way forks tend to be: any OpenAI-compatible endpoint works, local models included, so "Qwen" names the default configuration and the deal behind it, not a lock-in. The deal changed, though, and the tutorials haven't caught up: the generous Qwen OAuth free tier (2,000 requests a day, trimmed to 1,000, then 100) was discontinued on April 15, 2026. What replaced it is the Alibaba Cloud Coding Plan — a quota-based subscription (per-window, weekly, and monthly request caps) that also plugs into Claude Code, Cline, and Cursor, and gets each new Qwen coder model first. The model evidence is genuinely strong for open weights: Qwen3-Coder-Next (80B total, 3B active, Apache 2.0, February 2026) posts 70.6–71.3% on SWE-bench Verified across three different scaffolds in its technical report — a consistency result this whole article's thesis would predict matters more than any single peak. Qwen Code is not alone, either: it is the oldest and largest of a five-agent wave from Chinese labs (ZCode, MiniMax Code, Kimi Code, MiMo Code), most of them forks or rewrites. As one survey put it: the harness is increasingly commodity; the model deal behind it is the product.
Pi is the counter-argument to everything above, from Mario Zechner (libGDX) and now Earendil, Armin Ronacher's company. Where Claude Code's system prompt runs 7,000–10,000 tokens before you type a word, Pi's is under 1,000. Where the others ship feature grids, Pi ships four tools — read, write, edit, bash — a plain loop, and a TypeScript extension system: no built-in subagents, no plan mode; if you want a feature, you ask Pi to build it, and extensions hot-reload into a session tree you can branch. The philosophy is context control: Zechner's claim is that harnesses injecting things behind your back is how outputs degrade, and the receipts are not nothing — with only those four tools, Pi placed 8th on Terminal-Bench running Opus 4.5. Pi 1.0 landed October 1, 2026, including the year's most honest u-turn: MCP support, folded into a sandboxed Codemode layer after Zechner had publicly argued Pi didn't need it, with the long-running orchestration split into a separate Pi Durable substrate so the core stays minimal. MIT core, 58,000 stars, two dozen providers plus local models. Pi is not trying to win the platform war; it is trying to be the harness a single developer can hold in their head — and grow.
| Claude Code | OpenCode | Muse Code | DeepSeek Harness | Qwen Code | Pi | |
|---|---|---|---|---|---|---|
| Maker | Anthropic | Anomaly | Meta | DeepSeek | Alibaba (Qwen team) | Earendil (Zechner) |
| License | Proprietary | MIT | Proprietary (beta) | MIT | Apache 2.0 | MIT core |
| Models | Claude only | 75+ providers + local | Muse Spark 1.2 | 7 provider families | Qwen3-Coder default; any OpenAI-compatible endpoint | 24+ providers + local |
| The bet | The complete harness: skills, hooks, teams | Neutrality: any model, one TUI | Model co-trained with the harness | The loop itself is a plugin | A lab harness for open-weight coders; the model deal is the product | Minimal core you extend yourself |
| Signature | Agent Teams + Agent SDK | plan/build agents, LSP auto-load | Parallel sub-agents in git worktrees; replay-exact log | Modes; event-log resume/fork/replay | Gemini CLI lineage; Coding Plan gets new Qwen coders first | <1,000-token prompt; 4 tools |
| Price shape | Subscription or API | Free; BYO keys (Go $10/mo, Zen PAYG) | $1.25 / 1M input tokens | Free; BYO keys | Free software; Coding Plan subscription or BYO key (OAuth free tier ended Apr 2026) | Free; BYO keys |
| Evidence | Top TB 4.0 pair (64.85%, Oct 6); best scaffold in Particula's run | 212K stars; AA Coding Agent Index 53.6 (GLM-5.3) | Internal demos; independent runs pending | Minimal mode runs official evals; V4-Flash 82.7 TB 2.1 | ~27.6K stars; Qwen3-Coder-Next 70.6–71.3% SWE Verified across 3 scaffolds | TB top-8 with Opus 4.5, four tools only |
Choose by what you are actually buying. For a production default today, Claude Code has the deepest evidence and the most complete harness — you pay in lock-in. For sovereignty — your models, your keys, your local weights, no vendor able to revoke a login — OpenCode is the platform and Pi is the workbench; both are open-source, and both treat the model as a replaceable part, which is the correct long-term posture. Qwen Code makes the same point from the lab side: an Apache-2.0 harness whose value is the open-weight model deal behind it. Muse Code and DeepSeek Harness are the two to watch rather than adopt blind: Meta is running the price play with a co-trained stack, and DeepSeek is running the architecture play with a plugin runtime — and both labs have already shown they will post-train the model for the harness. That convergence is the real story of 2026. The harness started as packaging. Then it became the differentiator. Now the labs are training models and scaffolds as one artifact — which is, not coincidentally, the same loop Ornith runs in the open. The wars end when the harness builds itself.
Sources: Particula Tech scaffold comparison and Scale AI analysis, via dev.to and Digital Applied; Claude Code — mpiv.ai operator notes and MorphLLM (TB 4.0 board); OpenCode — Pinggy, Agent A Day (2.0), breach via Infosecurity Magazine; Muse Code — NeuralCoreTech and The Outpost; DeepSeek Harness — Automater Intel and CometAPI; Qwen Code — dosa.dev survey of Chinese-lab agents, OpenAIToolsHub review, Qwen3-Coder-Next technical report; Pi — Zechner's design notes, The Register (1.0/MCP), dev.to harness roundup.