Video stack
Animated explainers & social clips
A coding agent writes the scenes. Remotion previews and renders the video. Official skills supply the project conventions.
Claude Code / CodexRemotion + Skills
Some setup
THE AGENT HARNESS LIBRARY
Give your AI
a superpower.
Make videos. Build games.
Run a smarter business.
What is an agent harness? The software around a model that manages its tools, context, permissions and task loop—so it can act, check results and keep working.
Source-linkedTools + combinationsLast updated: 4 November 2026
START WITH THE OUTCOME
What do you want to do?
Useful combinations, with setup steps and honest limitations. These are researched recipes, not all hands-on tested.
11 workflowsCommunity-informed. Primary-source checked.
Video stack
Trims, crops & repeatable exports
Pair a coding harness with FFmpeg for scripted edits. Keep the original files and review a low-resolution preview before final export.
Coding agentFFmpeg
Technical
Games stack
Prototype → run → fix → play
Use a coding agent in your Godot project, then close the loop with engine runs and error feedback. Add a reviewed MCP bridge when you need editor tools.
Claude Code / CodexGodot + optional MCP
Some setup
Games stack
Scenes, scripts & test runs
Connect a compatible coding assistant to the community MCP for Unity bridge. The agent can call editor tools instead of only editing text.
Compatible coding agentMCP for UnityUnity Editor
Some setup
Business stack
Research → draft → human approval
n8n’s AI Agent node can choose tools inside an app workflow. Put human review on tools that send messages or change important records.
n8n AI AgentApp tools + review
Some setup
Business stack
Recurring briefs & reusable routines
Hermes combines an agent loop, tools, memory and reusable skills. Start with a narrow recurring task, then expand based on observed results.
Hermes AgentMemory + Skills
Some setup
Business stack
Scheduled research & private reports
OpenClaw supplies the assistant runtime, channels and scheduling. A restricted research workflow can collect sources and draft a daily briefing.
OpenClawSearch + approved tools
Technical
Coding stack
A frontend with a visual feedback loop
Pair a coding harness with Playwright CLI and skills, or MCP when you need an ongoing browser session. Inspect the local app with an isolated browser profile.
OpenCode / coding harnessPlaywright CLI + SkillsPlaywright MCP (alternative)
Some setup
Research stack
Long tasks, files & source trails
Deep Agents bundles planning, filesystem and context management. Add your own approved search tools and durable state for longer research tasks.
Deep AgentsSearch + state backend
Technical
Coding stack
A small core, your own workflow
Pi provides a compact coding-agent foundation. Add project instructions, selected extensions and a test loop rather than installing every possible integration.
PiProject tools + tests
Technical
Business stack
A task interface with a working agent
Use Codex SDKs for background jobs, or app-server for a product with ongoing conversations, streamed progress and approval controls.
Codex SDK / app-serverYour app and approved tools
Technical
THE MODEL IS ONLY THE STARTA brain needs
A brain needs
a way to act.
A model generates the next step. A harness runs the tools, manages the working context and feeds results back into the loop.
A useful stack adds the right domain tools—like a video renderer, game engine or business connector. A repository alone does not make a complete workflow.
01ModelReason about the task
02HarnessContext · tools · permissions · loop
03Real workAct → inspect → improve
A LIBRARY THAT KEEPS UP
Sources over hype.
We use current community discussions to find useful workflows, then check official documentation, repository ownership, licenses and release status. Every profile links its evidence.
1Discover2Verify3Publish
Daily research not yet verified
The current collection was refreshed editorially. Scheduled research has not published a verified refresh; failed attempts preserve the last published collection.
Agent harnesses, explained.
Is a harness the same as a model?
No. The model does the reasoning; the harness supplies the execution loop, tools, context management and permission boundaries around it. Different harnesses can use the same model.
Can a harness be a combination of tools?
Yes. This library separates a core harness from a composed workflow. Claude Code plus Remotion and its skills is a video stack; Remotion by itself is a renderer, not an agent harness.
Are these the “best” or all tested?
They are curated starting points, not a leaderboard. Community experience is anecdotal. Documentation checks establish what projects describe, not guaranteed performance or a security audit. Each workflow states its review level; this refresh did not include hands-on testing of the listed tools.
How is the library kept current?
Entries retain dated sources and change history. The isolated OpenClaw workflow is scheduled daily, with read-only research tools and a validator before publication. Its live status above reports whether a successful run has actually completed. Failed runs preserve the previous catalog.