Gareth Paul Jones
Product Lead, AI
Poe / Quora
Things I’ve learned
by making things.
Coding Agents with Swimlanes
Coding Agents with Swimlanes
Separate places for researching the task, running agents, and checking the result. The arrangement makes the work visible instead of burying it in one conversation.
Then move the task through plan → act → verify → QA → release.
See the actual walkthrough, from 3:14 ↗
A Multi-Signal Reward Function
A Multi-Signal Reward Function
Combining helpfulness, safety, and truthfulness signals. The session uses imperfect classifier proxies; it’s an experiment in feedback, not a validated scorecard.
Read the description and watch the experiment ↗
Evals for AI Agents
Evals for AI Agents
Testing agents using live environments, real tasks, and reproducible metrics. A session about what the system can actually do.
Watch the evaluation session ↗One from the older files.
In 2016, I wrote about Maskito, a face-filter app that failed. Adding login screens and more features had distracted me from the thing people actually wanted: better filters.
The retrospective: design, coding, distribution ↗