# Anyone else building their own agent harness instead of relying on out-of-the-shelf frameworks? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/aa7d27de8b8b7c48

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-20 14:54 ingested 2026-10-21 02:35

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xauanf/anyone_else_building_their_own_agent_harness/)

## retained text



```
Curious where everyone here lands on the "build vs. wrap" spectrum when it comes to running agent workflows for actual, daily work.

Between orchestration libraries like LangChain/LangGraph, off-the-shelf multi-agent frameworks, and native CLI tooling (Claude Code, Codex wrappers, etc.), the options are everywhere. Yet, it feels like half the builders I talk to eventually get frustrated by framework lock-in or context bloat and just end up scripting their own custom harness.

If you’ve rolled your own harness or bespoke workflow for a specific task:  
**1. What was the specific task or failure mode that pushed you to build your own?** (e.g., state recovery, tool sandboxing, deterministic linting/test gates, API token bleeding?)

**2. What does your stack actually look like?** Did you build on top of LangGraph/CrewAI, wrap an existing harness (Claude/Codex/Gemini CLI), or just write a bare-metal loop in Python/TypeScript with standard API calls?

**3. Where did it get messy?** Context window management, runaway tool loops, or overengineering a state machine that was supposed to "stay simple"?

Would love to hear what worked, what blew up, and whether you’d build it from scratch again or stick to managed tooling next time.
```



document fetched 2026-10-21 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xauanf/anyone_else_building_their_own_agent_harness
