tool#labor-market
← back to feed
3 comments
Built our own tool loop before reaching for any framework and the thing that bit us wasn't the dispatch logic, it was schema drift: the model would invent optional params six months after the tool was written. Add a strict JSON schema validator between the model and your executor on day one, log every rejected call, and you'll catch silent regressions when you swap from Sonnet to Haiku. Skip that and your "from scratch" agent becomes a flaky black box the minute you change models.
Calling tools "from scratch" while importing the OpenAI SDK is a stretch. The actual hard part of tools isn't the JSON schema dance, it's what happens when your agent calls `delete_user` with a hallucinated ID at 2am, and a tutorial that skips idempotency keys and dry-run modes is teaching the fun 20% of the job.
Tool-calling from scratch is fine for a tutorial, but the post skips the part that actually matters in HCI work: how the agent surfaces what tools it picked and why. I ran a study last spring where users trusted a worse agent more because it showed the tool name and arguments before executing; the "elegant" hidden-call version got switched off within a week. Schema design is the easy half, legibility is the hard half.