The agent harness scorecard: cowork and coding on two axes
A feature checklist cannot tell a kernel sandbox from a git worktree. Two axes can. We placed Kendr and 22 rivals on a depth-by-breadth map, across both competitions it fights: cowork and coding.
On a ten-dimension architecture-depth rubric as of mid-2026, the coding leaders are Claude Code and Kendr Code (88 each), then Codex (82), Cursor (80), the DeepSeek Harness (79), and Devin (78); the cowork leaders are Claude Cowork (83), Microsoft 365 Copilot (82), and Kendr Colleague (81). Depth is scored separately from feature breadth, so the two never collapse into one misleading number. Kendr sits above the trend line on both axes — the highest depth for its breadth — leading edit-transaction safety, permissions, and crash durability, while trailing on reach, hosted execution, and published benchmarks. These are architecture judgements against a published rubric, not benchmark results.
Why two axes instead of a checklist
Most agent comparisons are feature checklists, and a checklist cannot tell a kernel sandbox from a git worktree — it counts both as “isolation.” This scorecard uses two axes instead. Architecture depth is a ten-dimension weighted rubric, anchored so that a ten means a mechanism is enforced by the system rather than requested of the model, and better than the rest of the field ships. Feature breadth is a flat count of how many capabilities a harness actually ships. A tool can be high on one and middling on the other; that is not a contradiction, it is the finding.[1]
Kendr runs one agent loop under two capability registries — Colleague for knowledge work and Code for software — so it competes in two markets at once. This report scores both: nine cowork assistants and fourteen coding harnesses, each on its own adaptation of the rubric, and each placed on a depth-by-breadth map. It is Revision 3 of an earlier coding-only edition, and unlike that edition it re-scores the whole field rather than only Kendr’s own row.[2]
Two products this revision adds did not exist in the prior field, and both press directly on Kendr’s thesis: OpenWork, an open-source Claude Cowork alternative with the same local-first, bring-your-own-model, air-gapped positioning[3]; and the DeepSeek Harness, a composable coding framework whose append-only event log is the same state-first idea behind Kendr’s durability lead, and which can even run Claude Code and Codex as sub-agents.[4]
The coding rubric: ten dimensions
Each dimension is scored 0 to 10. The scale is anchored: 10 means the mechanism is implemented, enforced by the system rather than requested of the model, and better than what the rest of the field ships; 5 means a working but ordinary implementation; 3 or below means the concern is handled by convention or not at all.
| Code | Dimension | What a high score requires |
|---|---|---|
| Loop | Agent loop and control flow | A real tool-calling loop with mid-turn steering, interruption, read-only modes, runaway breakers, and conversation rewind |
| Edit | Edit engine and tool surface | Typed tools rather than raw shell; preconditioned writes; atomic multi-file transactions with rollback; checkpoints and undo; diff review |
| Ctx | Context engineering | Ignore-aware discovery, structural or semantic retrieval, an instruction-file hierarchy, size-aware compaction, and cross-session memory |
| Iso | Isolation and sandboxing | Kernel-level or container isolation, network egress control, checkout isolation, and an honest per-OS statement of what is contained |
| Perm | Permission and approval model | Granular gates, durable pending approvals, reusable allow rules, and identical enforcement for delegated sub-agents |
| Dur | Durability, recovery and audit | A durable invocation log as source of truth, single-writer leases, idempotent starts, restart reconciliation, replay, and tamper evidence |
| Model | Model layer | Provider breadth, bring-your-own-model, local inference, native tool-call parity across providers, and cost controls |
| Ext | Extensibility | MCP, skills, plugins, hooks, sub-agents, a published SDK or protocol, and cross-ecosystem interoperability |
| Surf | Surfaces and orchestration | Terminal, IDE, desktop, web and mobile reach; headless mode; parallel agents; worktree isolation |
| Cloud | Cloud, CI and delivery | Execution that continues with the laptop closed, hosted runners, scheduled cloud runs, and CI-native integration |
The coding scorecard
Fourteen coding harnesses, scored on the ten dimensions. Claude Code and Kendr Code tie at the top on depth, but for opposite reasons: Claude Code leads on surface reach, extensibility, and cloud delivery; Kendr Code leads on edit transactions, permissions, and durability. Devin, the DeepSeek Harness, and Jules are scored here for the first time.
Read these as architecture judgements, not benchmark results — no task-success rate feeds any total.
| Harness | Loop | Edit | Ctx | Iso | Perm | Dur | Model | Ext | Surf | Cloud | Total |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Code | 9 | 10 | 9 | 9 | 10 | 10 | 7 | 9 | 8 | 7 | 88 |
| Claude Code | 10 | 9 | 9 | 8 | 9 | 7 | 6 | 10 | 10 | 10 | 88 |
| OpenAI Codex | 9 | 9 | 8 | 10 | 8 | 7 | 6 | 8 | 8 | 9 | 82 |
| Cursor | 9 | 9 | 10 | 6 | 6 | 6 | 8 | 7 | 10 | 9 | 80 |
| DeepSeek Harness | 9 | 7 | 8 | 9 | 8 | 8 | 9 | 10 | 6 | 5 | 79 |
| Devin | 9 | 7 | 8 | 9 | 7 | 6 | 5 | 8 | 9 | 10 | 78 |
| GitHub Copilot | 8 | 7 | 7 | 7 | 7 | 6 | 8 | 8 | 9 | 10 | 77 |
| Gemini CLI | 8 | 8 | 8 | 7 | 7 | 6 | 5 | 8 | 8 | 9 | 74 |
| OpenHands | 8 | 7 | 7 | 9 | 6 | 6 | 8 | 8 | 7 | 6 | 72 |
| Windsurf | 8 | 8 | 9 | 5 | 6 | 6 | 7 | 7 | 8 | 7 | 71 |
| Cline | 9 | 8 | 7 | 5 | 8 | 5 | 10 | 8 | 6 | 4 | 70 |
| Kiro | 7 | 7 | 7 | 6 | 8 | 7 | 6 | 7 | 7 | 7 | 69 |
| Jules | 8 | 7 | 7 | 9 | 6 | 5 | 4 | 6 | 6 | 9 | 67 |
| Aider | 7 | 8 | 9 | 3 | 6 | 5 | 9 | 5 | 5 | 3 | 60 |
Coding: depth against breadth
Depth rewards doing hard things well; breadth rewards doing many things. Ranked by depth, with the breadth score alongside, the trade becomes visible. Claude Code leads both. Kendr Code matches it on depth at 88 but sits nine points lower on breadth — the clearest ‘above the trend line’ position in the field, because three field-leading dimensions carry weight that a flat feature count averages away.
Aider is the mirror image: a low breadth score (39) that undersells a world-class tree-sitter retrieval engine the depth rubric credits at 9. That divergence is the whole reason for scoring on two axes.
| # | Harness | Depth | Breadth |
|---|---|---|---|
| 1 | Claude Code | 88 | 86 |
| 2 | Kendr Code | 88 | 77 |
| 3 | OpenAI Codex | 82 | 76 |
| 4 | Cursor | 80 | 76 |
| 5 | DeepSeek Harness | 79 | 72 |
| 6 | Devin | 78 | 76 |
| 7 | GitHub Copilot | 77 | 65 |
| 8 | Gemini CLI | 74 | 49 |
| 9 | OpenHands | 72 | 68 |
| 10 | Windsurf | 71 | 54 |
| 11 | Cline | 70 | 52 |
| 12 | Kiro | 69 | 53 |
| 13 | Jules | 67 | 48 |
| 14 | Aider | 60 | 39 |
The cowork rubric: the same idea, adapted
Cowork assistants do office work — email, calendar, documents, research, browser action — so four coding-specific dimensions are replaced. Deliverables, knowledge work, connectors, and reach stand in for the edit engine, context engineering, isolation, and surfaces. The agent loop, permissions, durability, model layer, and extensibility carry across both rubrics unchanged.
| Code | Dimension | What a high score requires |
|---|---|---|
| Loop | Agent execution and autonomy | Multi-step planning, sub-agent fan-out, mid-task steering, and a defensible stopping contract |
| Deliver | Deliverables and artifacts | Real office files — spreadsheets, decks, documents — a versioned artifact library, and in-place revision |
| Know | Knowledge work | Multi-step deep research with citations, plus browser or computer use that can act, not just read |
| Connect | Connectors and integrations | Native email, calendar, files and business apps, plus custom connectors over MCP |
| Perm | Permission and approval | Granular gates, durable pending approvals, reusable allow rules, identical enforcement for sub-agents |
| Dur | Durability and audit | An authoritative invocation log, restart recovery, and a tamper-evident trail rather than a transcript |
| Model | Model and privacy | Bring-your-own-model, local inference, provider choice, and a local-first or on-device data path |
| Ext | Extensibility | Skills, plugins, persistent memory, hooks, and a portable capability format |
| Reach | Surfaces and reach | Desktop, web, mobile, chat-platform presence, and an in-browser surface |
| Cloud | Cloud and continuity | Work that continues with the machine off, scheduled and background runs, and hosted execution |
The cowork scorecard
Nine cowork assistants on the adapted rubric. Claude Cowork and Microsoft 365 Copilot lead on the strength of connector breadth and reach; Kendr Colleague places third, carried by the same substrate that leads coding — durable approvals, an authoritative log, bring-your-own-model and a local-first data path — and dragged by one dimension only.
OpenWork is the one rival sharing Kendr’s local-first, bring-your-own-model lane; it scores well on model and extensibility but thin on deliverables, because its coding and document work is inherited from an underlying engine rather than surfaced as its own.
| Product | Loop | Deliver | Know | Connect | Perm | Dur | Model | Ext | Reach | Cloud | Total |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Cowork | 9 | 9 | 8 | 10 | 8 | 7 | 4 | 9 | 10 | 9 | 83 |
| M365 Copilot | 8 | 9 | 9 | 10 | 7 | 8 | 5 | 8 | 10 | 8 | 82 |
| Colleague | 9 | 9 | 9 | 6 | 10 | 10 | 9 | 9 | 4 | 6 | 81 |
| ChatGPT Work | 9 | 8 | 9 | 9 | 6 | 6 | 4 | 7 | 9 | 9 | 76 |
| OpenWork | 7 | 6 | 6 | 9 | 6 | 5 | 10 | 9 | 7 | 6 | 71 |
| Gemini Workspace | 8 | 8 | 9 | 7 | 6 | 6 | 4 | 6 | 8 | 8 | 70 |
| Perplexity Comet | 8 | 8 | 9 | 5 | 5 | 5 | 6 | 5 | 9 | 8 | 68 |
| Manus | 9 | 9 | 9 | 5 | 6 | 5 | 4 | 4 | 7 | 8 | 66 |
| Notion AI | 7 | 6 | 6 | 8 | 5 | 6 | 6 | 7 | 7 | 7 | 65 |
Cowork: depth against breadth
The same two-axis view for cowork. Colleague ties Claude Cowork on breadth and lands third on depth — a strong position undercut by a single dimension, Reach, where it scores 4 for being desktop-only with no phone, web session, or chat-platform presence. None of the cloud incumbents share its local-first, offline, bring-your-own-model quadrant.
| # | Product | Depth | Breadth |
|---|---|---|---|
| 1 | Claude Cowork | 83 | 76 |
| 2 | M365 Copilot | 82 | 67 |
| 3 | Kendr Colleague | 81 | 76 |
| 4 | ChatGPT Work | 76 | 67 |
| 5 | OpenWork | 71 | 68 |
| 6 | Gemini Workspace | 70 | 55 |
| 7 | Perplexity Comet | 68 | 55 |
| 8 | Manus | 66 | 56 |
| 9 | Notion AI | 65 | 50 |
Where Kendr lands, including where it trails
On coding, Kendr Code ties Claude Code at the top of the depth axis (88) while sitting nine points lower on breadth. That comes from three dimensions it leads outright: a transactional edit engine that can roll a twelve-file change back as a unit, approvals that survive a process kill with their original arguments, and an invocation log treated as the source of truth rather than a transcript. Those are exactly the marks a breadth count averages away.
On cowork, Colleague lands third on depth (81), behind Claude Cowork (83) and Microsoft 365 Copilot (82), and ties Claude Cowork on breadth. One dimension drags it: Reach scores 4. Close that gap and the same substrate that leads coding would lead here too.
The honest gaps are the same on both sides, and they are stated on the rubric rather than hidden: no hosted plane for interactive runs, a model transport that degrades to a constrained JSON protocol under most of the catalogue, and — the loudest one — no measured benchmark on any subsystem yet.
Method, limits, and the full matrices
Nothing here is a benchmark. Depth scores are architecture judgements against a published rubric; no SWE-bench number, task-success rate, or latency figure feeds any total, and a weaker harness paired with a stronger model still wins most days. Breadth is a capability count, not a quality measure.
This is a dated snapshot — early to mid 2026 — and several products moved after the window closed. Model version strings churn fastest of all and were deliberately left out of the tables. The two axes are always shown together and never merged into a single number, because they answer different questions.
The full picture — all twenty-three harnesses on the depth-by-breadth map, and the exhaustive cowork and coding capability matrices, every feature and every product colour-coded — is in the interactive companion: the agent harness scorecard.
Frequently asked questions
Is this a benchmark like SWE-bench?
No. Every score is an architecture judgement against a published, anchored rubric — how the machine is built — not a measured task-success rate. No SWE-bench, latency, or output-quality figure feeds any total, and a weaker harness with a stronger model often wins in practice.
Why does Kendr rank first on depth but second on breadth in coding?
They measure different things. Depth weights ten architectural dimensions, where Kendr leads three outright — edit transactions, permissions, and durability. Breadth counts how many features ship, where Kendr posts deliberate zeros on reach, cloud, and a first-party model. The gap between the two is the breadth-for-depth trade Kendr has chosen.
What changed from the earlier coding comparison?
This is Revision 3. It adds the cowork field, re-scores every competitor rather than only Kendr, introduces a depth-by-breadth map so the two metrics never merge into one number, and adds three products the earlier edition missed: Devin, the DeepSeek Harness, and Jules.
How were cowork assistants scored differently from coding harnesses?
Cowork uses an adapted rubric. Deliverables, knowledge work, connectors, and reach replace the coding-specific edit-engine, context, isolation, and surfaces dimensions. The agent loop, permissions, durability, model layer, and extensibility carry across both.
Sources and evidence
Primary and authoritative sources used for factual claims. Company research and executive forecasts are labeled as such in the article.
- 1
- 2Coding agent harnesses compared: 50 tools scored on architectureKendr Research · 2026-08-14
- 3OpenWork — the open-source alternative to Claude CoworkOpenWork Labs · 2026
- 4DeepSeek Harness repositoryDeepSeek AI · 2026
- 5Get started with Claude CoworkAnthropic · 2026
- 6OpenAI Codex documentationOpenAI · 2026