Blog
Practical notes on building, operating, and improving AI agents for real teams.
The right model for every agent is rarely the biggest one
A practical way to match model cost, context, speed, and reasoning depth to the work an agent actually does.
Read postLatest posts
9 storiesFrom a Slack mention to finished work, without losing the thread
How channel routing, shared context, and durable sessions turn everyday messages into reliable agent work.
Give an agent enough access to finish the job, and nothing more
A team-first approach to connections, knowledge, skills, and the permissions agents inherit at runtime.
Search company knowledge before reaching for another API call
Why agents should begin with indexed company context, then use live provider tools only when the work needs them.
Every automation should leave behind a session you can inspect
Triggers are only the start. The useful record is the request, reasoning, tool calls, result, cost, and follow-up in one place.
Let agents use secrets without teaching them to reveal secrets
Environment variables can stay opaque while programs still receive the credentials they need inside a sandbox.
Agent memory works better when part of it looks like a database
Sheets give teams a shared, inspectable place for records that need to survive a chat and change over time.
The quiet economics of ending idle sandboxes quickly
Small runtime decisions compound across thousands of agent sessions. Here is how we think about the idle tail.
A good agent handoff answers three questions before work starts
Who owns the request, what context travels with it, and where the answer returns determine whether the workflow holds together.
What changes when you operate AI agents as a team resource
The unit of management shifts from isolated chats to shared agents, capabilities, knowledge, and observable runs.
Put the ideas to work