Most AI products treat intelligence like a microwave. You put something in, press a button, take it out. Useful and fast. But the microwave doesn't know your team, doesn't remember what you decided last week, doesn't push back when you're about to make a mistake.
We're past that moment. AI coding agents can write, review, and ship code. They hold context across a long conversation. They ask for clarification when a spec is ambiguous.
They're closer to collaborators with a narrow job, a lot of speed, and a constant need for clear scope and review.
So where do they live?
The tool metaphor shapes everything
When your mental model for an agent is "a tool," you design the workflow around that. You call it from a CLI. You copy the output into a doc. You share it in a thread. The agent stays invisible to the team.
Which is a waste, because the agent has context worth sharing. It knows what it was asked to do. It made decisions. It ran into constraints. It has opinions about what to try next. All of that dies in the terminal scrollback.
What teammate-first design looks like
In hilos, every agent is a channel member. It has an avatar (a squircle, where people get circles). It gets mentioned. It replies in the room where everyone can see.
When an agent opens a PR, it posts a report card: what it built, how to test it, what it left open, what worried it. The team reacts, redirects, approves. The agent ships the next iteration.
The person reviewing was in the conversation the whole time. The approval is part of the thread, not a separate ceremony.
The trust gradient
Trust isn't binary. You wouldn't hand a new colleague the production database on day one, but you'd let them build a component.
Agent trust works the same way. The default in hilos is PR-first: the agent opens a PR, and a human approves the merge from the report card. Cautious teams turn on the gate, so the agent proposes a diff and waits for approval before it pushes anything. Some teams keep agents read-only and use them for research and planning.
The workspace should reflect wherever you are on that gradient, and let it move as trust is earned.
Circles and squircles
The avatar shape matters more than it sounds.
Circles are people. Squircles are agents. When a channel has four engineers and six agents in it, you need to know who said what at a glance. The shape carries that.
The accountability that makes this work
The reason you can treat agents as teammates at all is the review gate. Every agent-authored change has a human who approved it, and the audit log records both.
If something goes wrong, you know exactly what was approved, by whom, and what the agent said it was doing. Not a log file nobody reads... a conversation that was visible the whole time.
The agent proposes. A person approves. Everything else is built to keep that true as teams speed up.