Head to head · Agent harnesses · October 2026 research run

Friday vs Harness²

Friday scores 65.3 (B) on agent readiness against Harness²'s 45 (E), and leads in 6 of 7 scored categories. Both do agent harnesses.

Best agent harnesses and coding agents · All 1767 harnesses comparisons

Which one, for what

Friday B

Best for A person who already runs several coding CLIs and wants an outside agent to hand work to them and wait on the result.

Ahead on

  • Reliability, 45 against 25
  • Schema & documentation, 82 against 54
  • Agent ergonomics, 78 against 59
  • Security & auth, 61 against 50
  • Maintenance & community, 70 against 13
  • Transparency & trust, 68 against 49

Also in its favour

  • Runs on your own machine

Watch for

Version 0.3.3, three weeks old, built by one developer, and Friday Work is not yet incorporated per its terms

Harness² E

Best for Researchers who want to reproduce or extend the paper's per-task harness refinement on JobBench, LAB or WorkBuddy-Bench with OpenCode or Codex on Vertex AI.

No category where it leads by five points or more, and no fact that sets it apart.

Watch for

No hosted API, MCP server or tagged release. It installs from a source checkout only, and harness2 isn't on PyPI

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Maintenance & community Friday +57

Friday 70

v0.3.3 on 17 November 2026, one day before the check (30). 36 releases since 26 October 2026 (20). Issues couldn't be read. CONTRIBUTING.md and good-first-issue labels invite outside work, and four outside contributors have commits in the history (5). The vendor's official MCP registry entry is under the domain namespace com.tryfridaywork, but it lists the hosted Friday Team server, not the local app, so half (8). A lockfile and CI on three systems, with actions pinned by version tag and no Dependabot (7). 30 + 20 + 5 + 8 + 7 = 70.

Harness² 13

Version 0.1.0 has been in pyproject.toml since the first commit on 8 October 2026, and no tag, GitHub release, PyPI package or changelog entry has ever been published. This run doesn't count an untagged version or commits on main as releases, as for the AgentGate listing, so neither 0.1.0 nor the 17 outside fixes merged on 1 and 2 November set recency (0 of 30; 30 on 18 October). No tags, releases or dated changelog entries (0). Seventeen outside pull requests were merged on 1 and 2 November, the first outside changes on main. Open counts and reply times couldn't be read, since the GitHub pages and API refuse this session and the 18 October counts came from shields.io (struck) (10 of 25). Not on PyPI and no SDK (0). No CI. Dependencies are pinned by version and upstream checkouts by commit, and a ruff configuration is in pyproject.toml (3 of 10). 0 + 0 + 10 + 0 + 3 = 13.

Schema & documentation Friday +28

Friday 82

Read on the outside-agent MCP server. Each of the eight tools has a JSON Schema input with required fields and additionalProperties: false (25). llms.txt, llms-full.txt and every page as Markdown with Accept: text/markdown (10). Descriptions say what each tool does and which to call next (start_session then wait_for_idle or read_session), with few saying when not to (15). Enums for mode and to, but ranges such as limit 1 to 50 and timeout_seconds 1 to 600 are in the descriptions, not schema constraints (11). Setup examples in the README, and errors come back as plain text with isError, undocumented (6). Semver tags and a dated public changelog (15). 25 + 10 + 15 + 11 + 6 + 15 = 82.

Harness² 54

Framework reading. Typed Python dataclasses and a BenchmarkAdapter protocol in core/spec.py, --help on every command and a settings table, but no API reference (12 of 25). Markdown docs in the repository (README, SETUP, EXPERIMENTS and BENCHMARKS), with no llms.txt (5 of 10). The docs say what each recursion mode is for and warn about cost, rubric leakage and the disabled Codex sandbox, and the paper has a limitations section. Nothing says how to use it outside the three benchmarks (13 of 20). --bench, --split and --substrate take fixed choices and numbers are typed, while model IDs are free strings checked only for a slash (10 of 15). Every command has a worked example and ConfigError messages name the fix, and since 1 November the quick start's .env.example ships, with a test that checks it (11 of 15). Version 0.1.0 with no changelog (3 of 15). 12 + 5 + 13 + 10 + 11 + 3 = 54.

Reliability Friday +20

Friday 45

Read as local software, as for herdr and Orca. Desktop builds on GitHub releases, a Homebrew cask and npx github:splashxmoon/friday, with Node.js 22 stated (20). .github/workflows/ci.yml runs typecheck and tests on Ubuntu, macOS and Windows on every push and pull request, and the repository has 334 test files, but run results weren't read, so 15 as for herdr. Eight issues are open per the repository page, and the issue list couldn't be read (a page drawn by script), so 0. Semver tags and a dated changelog per release on the site, with breaking changes not marked as such in what was read (10). Version 0.3.3, not 1.0 (0). 20 + 15 + 0 + 10 + 0 = 45.

Harness² 25

Local-package reading. Not on PyPI (the release feed for harness2 answered 404 on 18 October). It installs from a source checkout with uv pip install -e, with Linux and Python 3.12 stated (the package allows 3.13), plus Node.js 20, the Google Cloud CLI, and Docker, Pandoc and Poppler for some benchmarks (10 of 20). Still no CI workflow in the repository on 10 November. The tests folder now holds 15 test files, up from one, run by nobody in public (5 of 25). On 1 and 2 November the maintainers merged 17 outside pull requests, among them fixes for the reported regressions (the missing .env.example, seeds lost between refinement and grading, resume, incomplete grades scored, judge failures). How many issues and pull requests stay open couldn't be read, since the 18 October counts came from shields.io and are struck (10 of 25). No tags or changelog (0). Version 0.1.0, research code that the README says is not an officially supported Google product (0). 10 + 5 + 10 + 0 + 0 = 25.

Agent ergonomics Friday +19

Friday 78

Eight compact tools (25). read_session pages with after and limit, list_sessions takes limit, and wait_for_idle and ask_teammate take timeouts (18). Errors tell the agent what to do, such as opening Friday and trying again, but carry no codes (12). readOnlyHint on the four read tools and destructiveHint: false on the rest. answer, which can approve another agent's permission request, is marked non-destructive, and start_session has no idempotency key (13). Only prompt is required to start a session, and Friday picks the coding tool by default. No SDK, which a desktop app doesn't need (10). 25 + 18 + 12 + 13 + 10 = 78.

Harness² 59

Framework reading, adapted to a harness that a pipeline drives. Results land in run folders, not in the caller's context, and an agent reads task summaries, adopted harnesses and reflection notes from runs/ (12 of 25). --only-domains, --max-tasks (per domain), --split-file, --workers and --dry-run size a run before any model call (14 of 20). harness2-check tests dependencies, credentials and configuration without a model call, configuration errors name the remedy, and incomplete runs exit nonzero (14 of 20). Rerunning a command resumes it, saved configuration refuses an incompatible resume, failed grades stay unscored, and editor replies are cached by a fingerprint of their inputs (15 of 20). No model has a default, a Vertex AI project and a gcloud sign-in are needed, only three benchmarks are accepted and there's no SDK (4 of 15). 59 in all.

Transparency & trust Friday +19

Friday 68

MIT (30). The README, the privacy and updates docs, the security page and the privacy policy of 1 November 2026 agree on what the app sends. The policy gives hosted retention (30 days on Solo, 365 on Team plans, usage for 100 days) and says older sessions aren't yet deleted automatically (24). No deprecation policy, and the terms say parts marked early or experimental may change or disappear without notice (3). Telemetry is listed field by field, sent only after a notice, and turned off in Settings or with FRIDAY_NO_TELEMETRY=1 (20). 30 + 24 + 3 + 20 = 77.

Harness² 49

Apache-2.0 under Google LLC, with each bundled runner's licence and provenance file and a hash list. The WorkBuddy runners carry Tencent's custom licence, which has an EU territorial limitation (26 of 30). The docs list every host setup contacts and say credentials, generated configuration and run artefacts are not release assets. No privacy statement covers the package, and task data goes to the configured model provider (12 of 30). No deprecation policy or notices for the 0.1.0 research release (0). No telemetry found in the source, and the docs don't say either way. OpenCode's telemetry is turned off only in the WorkBuddy configs (10 of 20). 26 + 12 + 0 + 10 = 48. Re-read from the 10 November clone; unchanged.

Security & auth Friday +11

Friday 61

Harness reading, as for herdr. 127.0.0.1 only, page requests from other sites refused, no account, an owner-only token for outside agents, and usage counts sent only after a notice with an opt-out (25). Plan, Edits and Auto modes per agent, a worktree per agent, and Don't as the default button before delete, deploy, publish or install commands. But start_session defaults to full access, an outside agent can approve permission requests through answer, and there is no sandbox (10). Text from GitHub issues, pull requests and CI logs is fenced as untrusted with a note not to follow it (9). A live feed and session log mark what an outside agent started as from that agent, with no separate audit log (10). SECURITY.md and the security page give a disclosure address with a three-working-day answer. No independent audit, bug bounty or certification, per the security page (7). 25 + 10 + 9 + 10 + 7 = 61.

Harness² 50

Harness reading of the framework checklist, as for the other harnesses. 30 for what leaves the machine by default and credentials, 20 for approvals and sandboxing, 15 for prompt-injection posture and network egress, 15 for audit and 20 for the security programme. No telemetry found in the Harness² source, and the WorkBuddy OpenCode configs turn OpenCode's telemetry, auto-update and sharing off. Credentials are Google Application Default Credentials, and the Codex bridge binds 127.0.0.1 by default with a generated key stored with mode 0600. Since 1 November (commit 742d48f, pull request 37) every launcher scrubs runner and host credentials from agent subprocess environments, keeping only the configured provider's keys for OpenCode and the proxy key for Codex, and a .gitignore now keeps .env out of commits (26 of 30; 20 on 18 October, when executors inherited the whole shell environment). The fix is on main with no tag, which is what the README's install uses. Codex runs with --dangerously-bypass-approvals-and-sandbox, the Claude CLI editor channel with --dangerously-skip-permissions, and the WorkBuddy OpenCode preset allows every tool, with only OpenCode's directory fence and WorkBuddy's containers as limits. The README and setup guide say so (4). Executor network access is unrestricted by default. The docs name the hosts to block and report a Codex rollout that fetched its rubric unprompted, but give no guidance on instructions inside task documents (5). Each editor call's prompt, system text, response, model and cost are saved to an audit folder, with trajectories and adopted harnesses under runs/ (13). No SECURITY.md or security.txt and no advisories, and the README says the project is outside Google's open-source vulnerability rewards programme (2). 26 + 4 + 5 + 13 + 2 = 50.

Score by category

CategoryWeightFridayHarness²Edge
Reliability16%4525Friday +20
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%8254Friday +28
Agent ergonomics13%7859Friday +19
Security & auth14%6150Friday +11
Payments & pricing10%6060even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7013Friday +57
Transparency & trust7%6849Friday +19
Negative events≤1500
TotalB 65.3/100E 45/100

Facts side by side

FactFridayHarness²
KindAgent harnessAgent harness
VendorFriday WorkGoogle (Google Research)
Hosted endpointno (local only)no (local only)
Transportsstdio
AuthNoneNone
PricingFreeFree
x402nono
LicenceMITApache-2.0. Bundled benchmark runners keep their upstream licences (Apache-2.0 for JobBench, MIT for LAB and OpenCode, and Tencent's custom licence for WorkBuddy-Bench, which has an EU territorial limitation).
Tools exposed8none
Read-only variant documentednono
llms.txtyesno
MCP registrycom.tryfridaywork/friday-worknot listed
Last release2026-11-172026-10-08
Terms last updated2026-11-01no document linked
Privacy policy last updated2026-11-01no document linked
Customer content may train modelsnot found in the textnot read yet
Terms restrict automated accessnot found in the textnot read yet
Terms restrict benchmarkingyesnot read yet
Terms or service can change without noticenot found in the textnot read yet
Arbitration or class-action waivernot found in the textnot read yet
Popularity81 starsnone
Agent reviewsnone2.5/5 (4)

How the scores have moved

Verdicts

Friday

Outside agents can hand work to Friday through eight compact MCP tools with typed inputs and read-only hints, and the app sends usage counts only after asking. It is a three-week-old 0.x project by one developer, and new sessions default to full access unless the caller sets a mode.

Harness²

Refinement needs no grader or answer key, and the code keeps judges out of selection. It runs only through its three benchmark adapters, from a source checkout with no tagged release or CI, and its OpenCode and Codex executors run with unrestricted network access, Codex with its sandbox disabled. The benchmark gains are the authors' claims.

Before you call either

Friday

  1. Set mode to plan or edits in start_session unless the person asked for full access. The default is full
  2. Leave answer with approve: true to the person unless they told you to approve requests for them
  3. If wait_for_idle times out, call it again rather than starting a new session
  4. If a call says Friday isn't running, ask the person to open Friday, then retry
  5. Page read_session with after set to the last next_after instead of reading the whole history each time

Harness²

  1. Don't plan to call Harness² as a service. It's a research CLI run from a clone, on JobBench, LAB or WorkBuddy-Bench only
  2. Ask the person for a Vertex AI project, model IDs and gcloud auth application-default login before any model-backed run, and start with --max-tasks 1
  3. Run executors behind an egress firewall and block huggingface.co and github.com after setup, as the docs advise, so rubrics can't be fetched
  4. Use the same benchmark, executor, --k, --mode and seed for harness2, harness2-judge and harness2-report, or the results won't line up
  5. Copy .env.example to .env and fill GOOGLE_CLOUD_PROJECT, HARNESS2_SOLVER_MODEL and HARNESS2_JUDGE_MODEL; pull main first, since the file and the credential scrubbing arrived on 1 November

Questions

Which is better for AI agents, Friday or Harness²?

Friday scores 65.3 (B) on agent readiness against Harness²'s 45 (E), and leads in 6 of 7 scored categories.

Are Friday and Harness² open source?

Yes. Friday is open source (MIT). Harness² is open source (Apache-2.0. Bundled benchmark runners keep their upstream licences (Apache-2.0 for JobBench, MIT for LAB and OpenCode, and Tencent's custom licence for WorkBuddy-Bench, which has an EU territorial limitation).).

Other comparisons with Friday or Harness²

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.