06-reference

data engineering central omnigent meta harness

2026-08-31·reference·source: Data Engineering Central·by Data Engineering Central (Substack curator)
meta-harnessmulti-agentagentic-toolingpoliciesharness-engineeringwatch

Omnigent — a commercial meta-harness that sits above Claude Code, Copilot, OpenCode

Why this is in the vault

Hands-on install-and-poke review of Omnigent, a "meta-harness" that wraps multiple coding-agent CLIs (Claude Code, Copilot, OpenCode) behind one interface, policy layer, and web UI with a multi-agent orchestrator ("Polly"); direct evidence for the harness-engineering thesis that the control layer above the model is where the value now concentrates.

The core argument

The author frames the problem as tool sprawl: too many coding agents, too many /skills, MD files, plugins, MCP servers, each demanding a different setup, with no shared layer to apply consistent rules across them. Omnigent installs via curl | sh, auto-detects existing harnesses (Claude Code, Copilot), and layers three things on top: (1) a single interface/session view across harnesses with a local web UI (port 6767) for watching, commenting on, and editing agent sessions from a browser or phone; (2) Policies — named, reusable guardrails (Safety and Cost Control categories) that can require permission before bash commands or git push, either applied ad hoc via chat or referenced by YAML path, currently session-scoped only; (3) Polly — a multi-agent orchestrator that decomposes a task, delegates sub-tasks to different agent vendors in parallel, and does cross-vendor code review, demoed with a "plan a Rust/Python DataFusion+Arrow wrapper" prompt. The author's verdict: policies and Polly are the standout features; the remote-access-from-phone feature (via Railway deploy) failed on a missing DATABASE_URL and the author was relieved not to have to set it up, flagging real burnout/mental-health concern about agents running from bed/phone 24/7.

Mapping against Ray Data Co

Omnigent is a commercial, portable version of the control layer RDCO has hand-built bespoke: the founder-approved policy gate ([[feedback_automode_classifier_hard_gate]] — deploy/production-write denial as a hard gate), the brigade multi-station pattern (station-spec-author → station-test-author → station-code-author → station-critic, fresh-eyes review baked in), and cross-vendor composition already scoped in [[2026-07-25-multi-agent-claude-codex-grok-composition-patterns]]. Omnigent's "Polly" is functionally the same idea as RDCO's brigade fan-out plus [[verify-dispatch]]/[[verify-strategic-output]] critics, but packaged as a generic product feature instead of a project-specific skill stack. Two takeaways: (1) validates that RDCO built the right layer — a third party is now selling the exact primitive (session-scoped policy + multi-agent delegation) as its core value prop; (2) exposes a gap — RDCO's policies live scattered across CLAUDE.md hard rules and skill logic rather than a single named, reusable, cross-session policy object, which is closer to what Omnigent ships out of the box. Same disposition as [[2026-08-31-herdr-agent-runtime]] (filed same day, also a third-party layer over existing coding-agent CLIs): promising primitive, not adopt-now — the failed Railway/Postgres deploy and session-only policy scoping are exactly the early-stage rough edges that argue for watch-not-install, consistent with [[feedback_mcp_install_security_review_default]]'s bias toward reviewing before installing rather than reactively.

⚠️ Sponsorship

Newsletter is sponsored by Delta Lake ("Thanks to Delta for sponsoring this newsletter... Content like this would not be possible without their support"), disclosed at the top of the issue. Delta Lake is unrelated to Omnigent or the meta-harness content itself — this is a recurring per-issue sponsor slot for Data Engineering Central, not sponsorship of the Omnigent review. No indication the Omnigent coverage itself was paid or influenced; treat the product opinions as the author's own but note the author had not gotten the remote-access feature working at time of writing, so the review is partial/early.

Related