Chapter 06
Working with the agent
How to read and steer an agent as it works on your task. The agent runs in its own instance and streams its progress to the chat in real time.
The task detail view
Open a running task to see:
- Instance status bar — the current instance state (see Creating and running tasks).
- Workflow panel — a summary of which workflow phase the agent is in.
- Chat — the agent's live messages, tool calls, subagents, and any approval cards waiting on you.
Reading the agent's messages
The agent types out its plan and progress as it goes — what it found, what it's about to do, and why. Messages stream in as they're produced, so you can follow along in real time.
Tool calls & subagents
As the agent works, it takes actions — searching the code graph, reading files, running commands. These show up as tool calls in the chat.
The agent can also spawn a subagent for a focused sub-task — most commonly an "explore" subagent that queries the code graph to find things. Subagent activity appears in the chat too; you usually just read it, no action needed.
Interactive cards
When the agent needs you, it shows an interactive card in the chat and pauses until you respond. There are a few kinds:
Question prompts
The agent asks you to choose between options — for example, which of two approaches to take, or which module a change belongs in. Pick an option (or type your own) and the agent continues.
Approval prompts
The agent asks for a yes/no — most often design sign-off before it starts implementing. Approve to proceed, or request changes and the agent will revise its plan.
Proposal cards (commit / push / PR)
When the work is ready, the agent proposes git actions one at a time:
- Commit proposal — shows the files to stage and the commit message; approve to commit.
- Push proposal — approve to push the branch to your remote.
- PR proposal — approve to open the pull request (see Reviewing and merging PRs).
Each card lets you approve or reject; reject to send the agent back to revise.
important
Model behavior
Vyazen can use different models for different phases of the workflow — for example, a stronger model for design, a faster one for routine steps. You don't pick models per message; the agent and your instance's configuration handle the switches. Which models are available depends on your instance (see Account & admin if you're an admin).
Interrupting a turn
If the agent is heading the wrong way, you can interrupt the current turn rather than waiting it out. The task detail view has controls to stop or abandon the current run; the agent halts and you can redirect it in your next message or retry the task.