Cool. ๐ค Apparently agents are more careful (anecdotally) if they are told "you have 20 minutes; take your time."
Makes sense. I assume agents are trained to default to quickest (reward hacking). If you relax that constraint they are "free to do the right thing"*.
TIL you can just ask the agent to investigate a security prompt/notification.
I was alarmed that iTerm was automatically allowed to run in the background. Codex sorted that out fairly quickly that it was Mac OS doing something for me. Tsk tsk.
TIL that codex can manage root agents you started.
Start a few codex sessions and rename them with /rename. Use a namespace as convenience for yourself. i.e. {feature}/foo, {feature}/bar
Then you can ask a managing agent to work with {feature}/* agents I have started.
@skillleaderbd Yeah percentage is arbitrary, but it's a good starting point for agents. Best not to let them judge if it's good enough. I tend to ask them to 2 weeks worth of sessions, make scorers, and provide a score. That plus the half advisory is a strong steer.
Don't forget to clean your agent files.
Cut half of the skills, half of the AGENTS, and half of your config files.
Newer models need less steering. Oversteer them and they will perform worse. Plus it'll save you some $$$ and delay the 5h or weekly limit.
@VisiveAI The agent is very smart on researching things like this. Any security prompt I now resort to asking the agent to figure out what caused it so I can decide if I'll allow/deny.
@VisiveAI Communicating with voice to an agent is drastically different than communicating with text. Text, I seem to be more terse. Perhaps to a fault (too vague for the agent). When speaking with voice, I am elaborate. Likely the ease of having to forget about grammar.
@VisiveAI I'd claim that prevention is too costly compared to investing in detection, patching, and shipping.
Trace everything. I'm biased but @braintrust is my go to.
@sinatayebat Yes! If you have follow up questions I use the CLI and/or the Braintrust MCP. If you need to connect the dots back to your code you can use Github + Loop: https://t.co/QvZGXj3S51
Grateful for the chance to share how Braintrust Patterns surfaces recurring issues across agent traces at scale. Weโre starting to see how critical Brainstore is to turning those patterns into concrete improvements.
https://t.co/GY9GKvbPZH
Vote informed, not swayed by campaign noise.
Use agents to research your ballot: ask about judgesโ past rulings, who backs each proposition, and what else theyโve supported. Agents can estimate tax impact and how funds would be distributed.
Don't forget to fact-check!
Linear + Realtime voice is a game changer. I practically have my own executive assistant.
I ask what's due or overdue, then go through tickets one at a time. The agent leaves comments, creates follow-ups, marks tickets done, or moves due dates.
No UI. All hands-free.