Agent Harnesses Compared
A harness is the program wrapped around a model: its loop, tools, memory, permissions and where it runs. On Terminal-Bench the same model scores 9 to 14 points higher in a tuned harness, so picking the harness matters about as much as picking the model. As of November 2026, Claude Code leads on hard multi-file work, Codex wins on cost per task, Cursor owns the editor, and open-source harnesses such as OpenCode and Cline win on model freedom.
At a glance#
Harnesses split by where you work with them. The big vendors now ship the same agent in several places: Claude Code runs in the terminal, the IDE and the web, and Codex runs in the terminal, the cloud and on your phone.
The filled chips tie you to one maker's models. The dashed ones let you bring any model, which is the main thing open-source harnesses sell.
The contenders#
Fifteen harnesses matter in November 2026. Two big shifts this year: Google replaced Gemini CLI with the closed-source Antigravity CLI for individuals on Jun 18, and Roo Code shut down on May 15.
| Harness | Maker | Runs as | Models | Licence | Price per month | 2026 change worth knowing |
|---|---|---|---|---|---|---|
| Claude Code | Anthropic | CLI, IDE, desktop, web, mobile, Slack | Claude only; Opus by default | Proprietary | Pro $20, Max $100, Ultra $200 | Max 20x renamed Ultra on Oct 30, now 10x Pro weekly |
| Codex | OpenAI | CLI, IDE, desktop app, cloud, phone remote | OpenAI only; GPT-6 family | CLI is Apache-2.0 | Free, Go $8, Plus $20, Pro $100 to $500 | GPT-6 at DevDay (Sep 29); Pro $200 cut to 10x Plus |
| Cursor | Anysphere, owned by SpaceX | IDE, CLI, cloud agents | Composer 2.5, Grok, Claude, GPT, Gemini | Proprietary | Free, Plus $8, Premium $30, Ultra $200 | SpaceX deal closed Aug 14; Agent SDK GA Oct 20 |
| GitHub Copilot | GitHub | IDE, CLI, cloud coding agent | 30+ models; BYOK and local in CLI | Proprietary | Free, Pro $10, Pro+ $39, Max $100 | Usage-based AI Credits since Jun 1 |
| Antigravity | Desktop IDE, CLI, SDK | Gemini 3.8 Flash, Gemini 4; Claude on Ultra | Proprietary | Free weekly quota; AI Pro and Ultra | Replaced Gemini CLI for individuals Jun 18 | |
| OpenCode | Anomaly (ex-SST) | TUI, desktop, web | 75+ providers, local models | MIT, 225.7k stars | Free; Go $10 | Two critical security holes patched (Jan, Sep) |
| Cline | Cline | VS Code, JetBrains, CLI, desktop, SDK | Any provider, local models | Apache-2.0, 70.9k stars | Free plus model; ClinePass $9.99 | SDK and rebuilt CLI with agent teams (May 13) |
| Kilo Code | Kilo | VS Code, JetBrains, CLI, cloud | 500+ models | MIT, 27.8k stars | Free; Kilo Pass $19 to $199 | Rebuilt on the OpenCode core (Apr 2) |
| Aider | Aider-AI | Terminal | Almost any, local models | Apache-2.0, 49.8k stars | Free plus model | Releases have slowed; still on v0.86 |
| Pi | Mario Zechner, Earendil | Terminal, SDK | About 25 providers | MIT, 118.2k stars | Free plus model | Creator joined Earendil (Apr 8) |
| OpenHands | All Hands AI | Web (Agent Canvas), CLI, SDK, cloud | Any model | MIT, 92.6k stars | Free self-hosted; Teams $500 for 10 users | Agent Canvas (Jun 16); runs other agents over ACP |
| Factory Droid | Factory | CLI, desktop, SDK | GPT, Claude, Kimi, GLM; BYOK | Proprietary | Free with BYOK, Pro $20, Max $200 | Droid 1.0 with script mode (Oct 29) |
| Amp | Amp Frontier (ex-Sourcegraph) | CLI, web, desktop, mobile | Claude, GPT, GLM; BYOK | Proprietary | Free; Individual $20 | Dropped editor extensions (Mar 5) |
| Devin | Cognition | Cloud agent, Devin Desktop (ex-Windsurf) | SWE-2 | Proprietary | Free, Pro $20, Max $200 | Windsurf folded in; SWE-2 model (Sep 10) |
| Hermes Agent | Nous Research | CLI, messaging apps, desktop, cloud | Nous Portal, OpenRouter, any endpoint | MIT, 260k stars | Free plus model | Self-improving skills and memory; v0.22.5 (Oct 22) |
Stars were read on GitHub on Nov 13. Prices are list prices from each vendor's page; Antigravity publishes no numeric quotas.
Build-your-own SDKs. The Claude Agent SDK is Claude Code as a library (Python, TypeScript): the same tools, loop, hooks, subagents and MCP, billed at API token rates. The OpenAI Agents SDK gives handoffs, guardrails, tracing and sandboxed execution, still at version 0.x. OpenHands, Cline, Cursor and Factory also ship SDKs now.
Benchmarks#
The harness moves scores almost as much as the model does. Hand the same model a tuned harness instead of a plain loop and it gains 9 to 14 points on Terminal-Bench 2.0.
The plain runs come from Vals and the tuned ones from Snorkel's board. They are separate runs, so treat the gap as a direction, not a precise number. Among the tuned harnesses the spread is small: GPT-5.5 scores 84.7% in NexAU-AHE, 83.1% in Capy and 82.2% in Codex CLI.
The cleanest harness test holds the model fixed. Composio Bench runs every harness on GPT-6 Astra over 30 stateful tool-use tasks.
This measures tool use across apps, not coding depth, and Composio sells tool integrations, so read it alongside Terminal-Bench. It does show that lean, open harnesses (Cline, Hermes Agent) can beat the vendor ones when the model is the same.
Terminal-Bench moved to version 4.0 on Aug 28 and 4.1 on Nov 9, 2026, and the versions don't compare with each other. On 4.1 the order holds: Claude Code with Opus 65.8%, Claude Code with Sonnet 5.5 63.9%, Codex with GPT-6.1 Sol 60.9% (BenchLM). Claude Code is not only a Claude shell: Kimi K3.1 Pro run inside it scored 60.9% on 4.0, ahead of every Codex entry.
Where sources disagree. AIToolsReview quotes Anthropic's own 4.0 figures (Sonnet 5.5 70.6%, Opus 66.4%), above the board's 61.8% and 64.8%. Morph reports Sonnet 5.5 costing more per task than Opus at max effort, which looks odd but is what the board shows.
What has no score. Cursor, Copilot, Antigravity, OpenCode, Aider, Cline and Hermes Agent have no Terminal-Bench 4.x entry. Cursor's Composer 2 scored 61.7% on the older 2.0 (Firecrawl). No board compares harnesses on SWE-bench; those scores are per model.
Scorecard#
Every harness trades something away. These scores are Claude's judgment from the benchmarks and reviews in this tab, not a measurement: use them to see the shape of each tool, not as a precise ranking.
Read across a row for a tool's profile. Claude Code and Codex are the two deep, vendor-backed harnesses; the open-source ones (OpenCode, Aider, Cline, OpenHands, Hermes Agent) win on model choice and openness and lose on polish at the top end.
Strengths and limits#
Each harness has one thing it does best and one cost you pay for it.
| Harness | Best at | Falls short on |
|---|---|---|
| Claude Code | Hardest multi-file refactors; deepest extension system (hooks, skills, plugins, subagents, MCP); same engine on every surface | Heaviest token use (3 to 4x Codex on the same task, per community tests); Claude models only; unpublished usage limits users hit fast |
| Codex | Cost per task; one agent across CLI, desktop, cloud and phone; long autonomous runs (Goal mode) | OpenAI models only; front-end and UI work; takes liberties with literal instructions; the most complex limit system |
| Cursor | In-editor experience; widest model menu; cloud agents with their own desktop and browser | Usage bills are hard to predict; an editor first, so weaker for headless and CI work |
| GitHub Copilot | Issue to pull request inside GitHub; cheapest paid entry ($10); agents on every tier | Tuned for low-to-medium tasks; cloud sessions capped at 59 minutes and one repo; surprise overage bills |
| Antigravity | Free entry point; parallel agents with browser use; newest Gemini 4 models | Closed source; vague weekly quotas and lockouts; no feature parity with old Gemini CLI |
| OpenCode | Any model including local; polished TUI; headless server mode | Quality follows the model you bring; two critical security holes this year; no Claude subscription login |
| Cline / Kilo Code | Open-source agent inside VS Code with any model; tops the one same-model harness test | You pay raw API prices; Kilo's rebuild drew complaints (forced worktrees, own terminal) |
| Aider | Small, precise, git-committed edits; cheapest way to pair with a strong model | Least autonomous; terminal only; releases have stalled at v0.86 |
| Pi | Minimal 4-tool core you can read and extend; branchable sessions | No MCP, subagents or plan mode out of the box; new contributors' PRs auto-closed |
| OpenHands | Self-hosted, model-agnostic platform; orchestrates other agents over ACP; scheduled automations | Real infrastructure work to run; governance sits in a $500/month Teams tier |
| Factory Droid | Script mode batches tool calls into a program, cutting tokens on long repetitive jobs | Gains shrink on short sessions and debugging; its benchmark wins are self-reported |
| Devin | Fire-and-forget tickets in parallel cloud VMs | Uneven on vague or open-ended work; own model only |
| Hermes Agent | Always-on personal agent: messaging apps, memory, schedules, self-written skills | Not a coding specialist; Windows and gateway bugs; opaque model errors |
Where one beats another. Claude Code beats Codex on hard refactors and extensibility; Codex beats Claude Code on cost, speed and mobile hand-off. Cursor beats both inside an editor. Cline beats Claude Code by 7 points when both run the same model on Composio Bench, so a harness tuned for one vendor is not automatically the best home for another vendor's model.
Head to head#
The matchups people actually decide between, with the winner on the question that usually settles it.
| Matchup | Winner | Why | Pick the other when |
|---|---|---|---|
| Claude Code vs Codex | Claude Code on quality | Leads Terminal-Bench 4.0 and 4.1 by 4 to 7 points; deeper hooks and plugins | Budget or volume matters: Codex runs a top task for $1.92 against $14.30 |
| Claude Code vs Cursor | Depends on where you sit | Claude Code for terminal, CI and long runs; Cursor for reviewing diffs in an editor | You want many vendors' models in one place: Cursor |
| Codex vs Cursor | Codex on autonomy | One agent on CLI, cloud and phone; cheaper per task | You live in the editor and want Composer's speed |
| Copilot vs Cursor | Cursor on agent quality | Stronger in-editor agent and cloud agents | The team runs on GitHub issues and Actions, or wants the $10 plan |
| OpenCode vs Claude Code | Claude Code on results | Tuned harness plus subscription pricing | You need local or open-weight models, or MIT code you can audit |
| Cline vs Kilo Code | Cline on stability | Mature, tops the same-model test | You want 500+ models through one gateway and parallel worktrees |
| Aider vs OpenCode | OpenCode on autonomy | Agent modes, subagents, server mode | You want tight, reviewable, auto-committed edits |
| OpenHands vs Agent SDKs | SDKs for products | Claude and OpenAI SDKs give a production loop in a few lines | You must self-host and stay model-agnostic |
| Hermes Agent vs Claude Code | Claude Code for code | Built for codebases | The job is a personal agent on Telegram or WhatsApp that remembers and runs on a schedule |
The benchmark gap between the top harnesses is now smaller than the gap between a tuned harness and a plain one. Choose on workflow, cost and model access first.
Pick this when#
Start from where you work and what you need back, then pick. Many teams run two: one deep harness for hard work and one cheap one for volume.
A common pairing in late 2026 is Claude Code for the hard 20% of tasks and Codex for the cheap 80%, with Cursor for anyone who reviews in an editor.
Sources#
Read between Nov 11 and Nov 13, 2026. Scorecard numbers are Claude-estimated from these pages; every other number comes from the page named.
Benchmarks
Terminal-Bench 2.0 leaderboard, Snorkel: best harness per model, updated Nov 9, 2026
Terminal-Bench 2.0 on Terminus 2, Vals AI: the same models in a plain harness
Terminal-Bench 4.0 and 4.1, BenchLM: harness and model pairs with cost per task
Composio Bench: seven harnesses on one model, run Oct 15, 2026
Comparisons and reviews
Vendor pages
Claude Code docs and Claude pricing
Cursor pricing and changelog
Google Developers Blog: the Gemini CLI to Antigravity CLI transition
GitHub repositories for OpenCode, Cline, Kilo Code, Aider, Pi, OpenHands and Hermes Agent: licences and star counts
Where sources disagree
Claude Code on Terminal-Bench 4.0: vendor-reported figures differ from the public board's. This tab uses the board.
Copilot Pro price: most pages give $10 a month and some give $19. This tab uses $10.
Gemini CLI cutoff: some reviews treat the May 19 announcement as the end date. Google's post puts the end of individual access at June 18.
Windsurf: lists disagree on whether it still counts as a separate product, so it is left out here.
OpenCode stars: pages read on different days give different counts. This tab uses the repository's own count.