RocketRide Office Hours: Better Model or Better Harness?
In February, LangChain took their coding agent from outside the top 30 to the top 5 on Terminal Bench 2.0 without touching the model. All they changed was what’s around it: a checklist before it says it’s done, a map of its working folder, a nudge when it goes in circles. That surrounding layer has a name now: the harness.
Here’s the catch. Most teams aren’t running one agent. They’re running a coding agent, a support bot, a judge on a different model, retrieval and a human review step. Each one has its own harness, with its own logs, its own evals and its own limits. Fix a mistake in one, and the others never learn.
Part 5 of our fortnightly agentic building series starts with what makes a good harness and ends one level up, with the meta-harness: a single layer that keeps every agent and model in line.
What we’ll cover
Why the same model can score very differently depending on what’s around it
Two habits from Mitchell Hashimoto, OpenAI and Anthropic: turn every mistake into a rule, and never let an agent grade its own homework
What happens when you’re running five harnesses instead of one
The meta-harness: shared traces, evals and limits across every agent and model, with RocketRide and our own support bot Ralph as the example
How many harnesses are you running, and what do they actually share?
About 10 minutes on the topic, then open discussion. Voice or chat, both count. Recorded, with a written summary after. You don’t need to have been at earlier sessions.
Join here: Join Discord Server