Is it worth building your own agent harness instead of using an existing one?
Someone spent about three months building a local agent harness that runs on their Mac and works with keys from several model providers, because they prefer keeping data local. Did others build their own harness too, and when is that worth the time compared to using an existing tool?
For most people, no. Existing harnesses already handle the hard plumbing: the tool-call loop, file edits, permissions, context compaction and retries. Building that again teaches a lot, but it is months of work before you get back to the job you wanted the agent for.
Building your own makes sense when you have a hard requirement the tools cannot meet, such as data that must never leave the machine, a model provider they do not support, or a workflow that does not look like chat plus tools. Write that requirement down first; if you cannot name one, extend an existing tool.
Most of the gains people credit to their custom harness come from things you can add to any harness: a clear rules file, a well-curated context, splitting independent work across parallel agents, and a habit of checking output before accepting it.
If you do build one, keep it thin. Let the model and the tools do the work, and spend your effort on logging every step so you can see why a run went wrong.
Listings mentioned
- Context engineering · skill by addyosmaniSets up rules files and context for a project, and helps when agent output quality drops mid-session.
- Dispatching parallel agents · skill by obraSplits two or more independent tasks across parallel agents without building your own orchestration.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.