TLDR: I did a full evaluation of omnigent as a fleet harness for my agent setup. The verdict: don't adopt. Two of its three headline features I already had. The win was the inventory, not the framework.

The Itch

Apollo (my personal AI agent, running inside Claude Code) has been growing into a real fleet.

CEO briefing daemon firing at 8 AM. Fathom transcript ingest. RAG reindex. An iMessage listener that catches me anywhere. A dashboard with Slack threaded in.

The operational pain was real. I'd watched Claude Opus silently burn on a dashboard restart — uncapped, unchecked, until I noticed the bill. I'd had the 8 AM briefing hang without a single error log. I'd hit TCC popup lockouts on headless claude -p processes that just… froze.

So when I found omnigent (github.com/omnigent-ai/omnigent, an open-source agent-orchestration framework — unified control plane over Claude Code, Codex, Cursor, custom agents, browser UI, mobile notifications, real-time session sharing), my instinct said: this is exactly what I need.

What I Almost Did

The pitch was compelling. Supervise multiple agents in one session. Agents that review each other. Cross-device control and OS push notifications. A governance layer: approval gates, spend governors, policy enforcement.

I was THIS close to migrating the whole fleet.

The Inventory That Stopped Me

Instead of building, I did something I should do more often: I mapped what I already had before deciding what to buy.

I went through every governance capability omnigent was selling. Here's what the inventory showed:

  • Approval / policy gate — Already have it, in its strongest form. Hard tool allowlist at the claude -p harness. --permission-mode dontAsk means deny-by-default — a non-allowed tool call aborts cleanly in headless mode. Plus iMessage Apollo's draft-confirm flow for anything going to other people. Omnigent would only centralize this. It wouldn't add capability.

  • Spend governor — Half built. Per-path caps exist: the claude-code-router shim, an OpenRouter hard cap, a Telnyx daily cap, Western-providers-only routing for daemons. But no unified cap across every claude -p + Agent-SDK call. The live hole: the Apollo Dashboard auto-rerank burns Agent-SDK credit on every restart, completely uncapped. That's real, and omnigent's governor would help — but it's also fixable with two lines of config I already control.

  • Remote fleet control — Genuinely missing. No single remote pane of glass. No way to list running agents from my phone, kill a hung daemon, or approve a pending action when I'm not at my desk.

Two of three headline features: already owned.

Why the Shape Is Wrong

Even if omnigent filled the gaps, my fleet is the wrong shape for it.

Omnigent optimizes for attended sessions — a human watching multiple agents work together in real time. My fleet is the opposite: unattended launchd cron daemons, hardcoded absolute paths, TCC/FDA grants, ~/.config/apollo secrets, per-daemon model routing. Rehosting that into a framework that's never seen it is a HUGE surface for the exact silent-fail traps I've already burned myself on.

And it's pre-1.0. As a single point of failure for every agent in the fleet — that's trading vendor lock-in for framework lock-in. I've written about the supply-chain-cooldown instinct before. This triggered it hard.

What I'm Actually Going to Build

The one real gap — remote fleet control — has two cheaper paths that don't require a framework adoption:

  1. Extend iMessage Apollo to list/kill/approve fleet agents (I'm already talking to it from anywhere)
  2. Expose the Apollo Dashboard behind Cloudflare Access (same pattern I used for Campfire)

Neither requires omnigent. Both are surfaces I already own.

Why This Matters to Me

The best decision I made this week was the code I didn't write.

I came in with an itch and a shiny new tool. The discipline was slowing down long enough to actually inventory what I had — and being honest that the instinct was right (fleet management does matter) while the proposed solution was wrong (a young framework as a choke point for everything).

If you're running your own agent fleet and feeling that same "I need a harness" itch: do the inventory first. Write down the three things the framework claims to give you. Then check whether you already have two of them. You probably do.

P.S. If omnigent ever hits 1.0 and I want to try it anywhere, the pilot is the autonomous overnight council builds — multi-agent, spend-heavy, sandbox-friendly, and isolated enough that a young framework can't take down the daemons if it falls over.