Monospace · AgentsNº 006 · Playbook Edition

The Harness
Above the Harnesses

Omnigent is Databricks' bid to become the orchestration layer above coding agents — one YAML spec driving Claude Code, Codex, Cursor, and Pi behind a governance engine that agents cannot talk their way around.

FLEET GOVERNANCE
omnigent-ai/omnigent · Apache-2.0 · DatabricksBenched 0.2.0 (June) + 0.4.0 (July 2026)
The Premise02

Not another coding agent — the layer that governs all of them.

A meta-harness: your agent spec contains zero harness-specific code, one field swaps the backend, and every action passes a policy engine that gates ALLOW, ASK, or DENY — with rules agents cannot weaken from the inside. Whether that survives contact with the code is what the bench keeps testing.

568klines of code
13 daysold at first bench
14harness families
24policy handlers
Pedigree: Spark · MLflow · DeltaThe Harness Above the Harnesses
The Machine03

Four layers, one session.

ClientsCLI · mobile-first web UI · native macOS — one session, three doors
ServerFastAPI + Postgres · 58 API paths · SSE + WebSockets · argon2id / JWT / OIDC / RBAC
Policy engine24 handlers · 6 phases · stricter-wins · DENY short-circuits
Meta-harnessSDK executors in-process · native executors drive real vendor CLIs behind a secretless credential proxy
One agent YAML selects the backend with one fieldThe Harness Above the Harnesses
The Bench04 · Why the plays are believable

Benched twice. Deterministic. Zero tokens.

June: five model-free trials against the pinned 0.2.0 wheel. July 4: the same suite re-run against 0.4.0, three runs per trial, stdout byte-identical — and our June scripts needed zero API adaptation two releases later. Every recipe is committed and re-runnable.

7.15sempty dir → version banner
3steps to first output
19/19policy checks, ×3 runs
0tokens spent
457 MBvenv · 100 deps · py3.12+
What we did not run: live multi-vendor sessions — tiers say soThe Harness Above the Harnesses
Play 01Verified A Weekend

Run a mixed fleet from one spec.

The scenario → the move

Your team already pays for two or three coding agents, and every automation script is welded to one of them. Write the agent once; swap the backend by editing a single field.

agent.yaml
executor:
  config:
    harness: codex-native   # ← the only line that changes
The receipt

Identical agent body validated on 17 of 18 harness ids on 0.4.0 — up from 11/18 in June. Allowlist grew 16 → 39 ids (~14 vendor families, incl. unadvertised kimi + kiro). ×3 runs, byte-identical.

Boundary — the one id that still rejects is databricks: their own.

Every June "wait-list" harness shipped within two releasesThe Harness Above the Harnesses
Play 02Verified An Afternoon

Give every agent a spending referee.

The scenario → the move

An autonomous agent with your API key and no dollar cap is a blank check with enthusiasm. Declare the budget once; every action from every harness is gated ALLOW, ASK, or DENY before it runs.

policies:
  - type: cost_budget
    hard_limit_usd: 25
  - type: ask_on_os_tools
The receipt

19/19 checks ×3 byte-identical runs on 0.4.0: over-budget DENY short-circuits · unknown model fails closed · DENY beats an ASK declared first · agents can't add policies without an unconditional ASK. New: per-subagent budgets.

Boundary — policies gate actions, not the quality of them.

The best-evidenced piece of the system, across two releasesThe Harness Above the Harnesses
Play 03Demonstrated A Weekend

Have one vendor review another's diff.

The scenario → the move

Cross-vendor review is the cheapest second opinion in coding-agent work — different model, different blind spots. The flagship Polly orchestrator delegates across vendors in parallel git worktrees, then routes the diff to a rival.

uv pip install omnigent==0.4.0
omni run examples/polly/
The receipt

The shipped example parses, validates, and loads under the pinned release. We did not run a live multi-vendor session — no vendor credentials on this bench — so orchestration quality is unmeasured here. That is what DEMONSTRATED means.

Boundary — Polly isolates with worktrees + a blast-radius policy, not the OS sandbox.

Steal the pattern even if you never adopt the serverThe Harness Above the Harnesses
Play 04Plausible A Sprint

Carry a session from laptop to phone.

The scenario

The overnight run finishes at breakfast; you approve the diff from the kitchen. One session reachable from CLI, browser, and phone over an append-only event log — with RBAC share, fork, and co-drive built in.

The receipt

Architecture read: 58 API paths · SSE + WebSockets · argon2id / JWT / OIDC · deterministic resume from the append-only log. Not exercised end-to-end on our bench — priced accordingly.

Boundary — real operational weight: Postgres, migrations, TLS, and one starlette CVE to patch before public exposure.

The architecture's promissory note — well-built, unexercisedThe Harness Above the Harnesses
Play 05Plausible An Afternoon

Hand your agents synthetic keys.

The scenario

A prompt-injected agent with real credentials is a breach; one with placeholders is an incident report. Native executors run vendor CLIs behind a secretless credential proxy — child agents only ever see synthetic stand-ins.

The receipt

Source read + security scan confirm the design: placeholder injection at the proxy, real secrets never enter the child process. Not exercised end-to-end on this bench.

Boundary — applies to the native-executor path; SDK executors follow the vendor SDK's own auth.

Governance for secrets, not just for spendThe Harness Above the Harnesses
The Fine Print10

No backdoors. The risk is operational, not malicious.

Full report: security scan in the study archiveThe Harness Above the Harnesses
The Ledger11 · The forwardable summary

Five plays, priced by evidence.

PlayTierEffortReceipt
01 Run a mixed fleet from one specVerifiedA WEEKEND17/18 harnesses on 0.4.0 (was 11/18), ×3 byte-identical
02 Give every agent a spending refereeVerifiedAN AFTERNOON19/19 checks ×3 · fails closed · DENY beats first ASK
03 One vendor reviews another's diffDemonstratedA WEEKENDPolly validates + loads under pinned release; live run untested
04 Session from laptop to phonePlausibleA SPRINT58 API paths · SSE/WS · auth read; not exercised
05 Hand your agents synthetic keysPlausibleAN AFTERNOONCredential-proxy design confirmed in source + scan
Tiers: Verified — we ran it · Demonstrated — its own example ran · Plausible — read, not runThe Harness Above the Harnesses
The Kicker12

Best for running your agents as one governed fleet.

If you run one play tonight, run the referee — it's an afternoon of YAML, it survived every probe across two releases, and it's the piece the rest of the fleet stands on.

uv venv --python 3.12 && uv pip install omnigent==0.4.0
omni --version   # 7.15 seconds from an empty directory to here
MONOSPACE · monospacemag.comField-tested reviews of trending repositories
← → / space · F fullscreen