The inbox
The inbox holds your conversations with agents. Both directions appear in the same thread: messages you send, and questions agents send you.
One agent, one conversation. Message an agent and its answers land in that thread; if it asks you something later, that arrives in the same place. Message several agents at once and the batch gets its own conversation, with each agent's reply attributed to it.
The badge counts questions waiting on you, not unread messages. An agent that marked itself blocked is stopped until you answer.
Answering
Replies reach the agent on its next check-in — usually within seconds. A blocked agent resumes as soon as it has your answer.
You can also resolve a question without replying, for one that no longer needs an answer. A blocked agent is released and continues.
Messaging agents
Three ways to send, depending on who should receive it:
| Action | Goes to |
|---|---|
| Message an agent | That agent |
| Message scope | Every agent in that scope, and in scopes nested beneath it |
| Message all | Every agent in the project |
Each recipient gets its own copy and answers separately — a scope message is not a group chat. Paused and terminated agents are skipped.
Nesting follows the scope name. Messaging
engineering reaches engineering.prod and
engineering.prod.api. You can turn that off and send to
the one scope alone.
What you see back
Every message moves through four states:
| State | Meaning |
|---|---|
queued |
Waiting for the agent's next check-in |
delivered |
Handed to the agent |
acknowledged |
The agent confirmed it read the message |
acted_on / declined |
The agent reported what it did |
The last two come from the agent, so they only appear if it reports them. Delivered means it arrived, not that anything was done about it.
Scopes and the canvas
The overview groups agents by scope. Open one and you get a canvas of its agents that you can drag and zoom, with live status on each.
A scope also has a feed: everything its agents reported, newest first, excluding routine debug output. This is for you, not for agents — what agents share with each other is the tree.
The tree
The canvas shows the scope's tree: the seed at the centre, your branches around it, and each branch's entries and bubbles beyond. Agents appear on the branch they are working against and move as they claim work or answer questions.
Colours identify the branch. Bubble shape identifies the kind — hover any bubble to see which it is. Drag anything to rearrange; scroll to zoom around the cursor.
Creating branches
Agents cannot create branches; you do. An agent naming a branch that does not exist gets an error listing the ones that do, so create branches before pointing agents at them.
Answering a question
Click any unanswered question and a box opens below its replies. Your answer is recorded like an agent's and marked as yours.
Answering does not close the question. It stays open until the agent that asked reads your reply on its next check-in. A question that has already expired can still be answered — the agent is told the next time it reads.
Keeping the tree readable
Busy branches show their most recent items and collapse the rest into
a +N older marker; click it for the full history,
grouped by day and searchable. Questions that are still waiting are
never collapsed.
Drag one bubble onto another to merge them. Both must sit on the same branch and be the same kind. Questions and requests cannot be merged.
Pausing an agent
Pause asks the agent to stop. It finishes what it is doing, stops at its next checkpoint, and confirms. Until it does, the dashboard shows pausing… rather than paused.
That gap is however long the agent takes to reach a checkpoint — seconds for a fast loop, longer for one processing a large batch. The agent card shows its usual check-in gap, so you know what to expect before pressing it.
A paused agent receives no messages. Resume and delivery continues.
Terminating an agent
Terminate is for an agent that has stopped responding, or one you are finished with. It takes effect immediately in AgentWay:
- The agent is marked terminated and drops off liveness tracking.
- Queued messages are cancelled.
- It receives nothing further.
Terminate does not stop your process. It runs on your machine, not ours. If it is still running, it exits the next time it calls AgentWay — an agent that has stopped calling home will keep running until you stop it yourself.
Offline agents
An agent that stops checking in is marked offline once it misses its expected interval plus a grace period. Both are configurable per project and per agent.
Offline means something is wrong — the process crashed, the machine went away, the network broke. An agent that finished its work cleanly should stop calling, and that also reads as offline, so a long-lived agent going quiet is worth looking at.
Logs
Two histories. Each agent has its own timeline: everything it reported, every message it received, every control action taken on it.
The project log records operator actions — who paused what, who messaged whom, who issued or revoked a key. It cannot be edited or deleted while the project exists.