flâneur

Building a harness for open models | writings

parthsareen.com · 3,537 words · saved by 1 readers

What open models expose about context, tools, policy, and the agent loop around them.

A model that “forgot” the task may have been shown mostly stdout. A model that uses the wrong tools may have been given a transcript where compaction rewrote the goal into something vague. And a model that seems “slow” may be running in a harness that adds the current time to the system prompt on every turn, breaking cache reuse. I know that last one a little too well. Most popular agent harnesses were built around closed frontier models. Closed models have huge context windows, a lot of recovery ability, and a pretty generous token budget, so a lot of harness bugs never surface. You can…

saved by

related reading