Chat
Chat is where you talk to your agents and swarms, see their work, and explore the results. This page covers starting a conversation, the two execution modes, pinning a specialist, attachments, and the rich panels that show an agent's thinking, sources, and files.
Starting a conversation
From the Chat area (or the chat icon in the dock), start a new conversation by choosing who to talk to. AlphaAgent shows your active agents and swarms together in one searchable list, each tagged by type. Pick one and the conversation opens.
Your recent conversations appear in the sidebar so you can jump back into them. See Getting started.
Execution modes
A toggle in the composer chooses how the agent works on your message:
- Express — direct execution for the fastest turnaround. No planning step and no knowledge-graph traversal. Best for straightforward questions and quick tasks.
- Deep Traversal — best for complex requests. The agent first builds a step-by-step plan, shows it to you, and waits for you to confirm before running. It can draw on a wired Knowledge Graph and runs up to 4 sequential steps.
Deep Traversal currently runs for single agents. When you chat with a swarm, the conversation is locked to Express and the Deep Traversal option is shown as coming soon. Swarm Deep Traversal is on the roadmap; until it ships, use Express for swarms (or run Deep Traversal against a single agent).
How Deep Traversal works
- You send your request in Deep Traversal mode.
- The agent responds with a plan, presented as a message.
- You review it and click Proceed to run it (or refine your request and re-plan).
- The agent executes the plan, streaming its progress.
Deep Traversal needs a Knowledge Graph
Every Deep Traversal flow starts from AMPG grounding against a Knowledge Graph — that grounding is what the planning and traversal steps build on. An agent that doesn't have a Knowledge Graph wired into its configuration therefore can't run Deep Traversal: there is nothing to ground against, so the flow can't begin.
If you try Deep Traversal on an agent without a Knowledge Graph, use Express mode instead (which doesn't require one), or wire a Knowledge Graph into the agent's configuration first. See Knowledge Graphs (AMPG).
If a request is too big for a chat plan, the agent explains that chat plans support up to four sequential steps and points you to Workflows for longer or parallel work. You can also convert a Deep Traversal plan into a draft workflow directly from the plan message — a handy way to turn a one-off plan into something repeatable.
Pinning a specialist with /direct
When chatting with a swarm in Express mode, you can send a single message straight to a specific specialist instead of letting the supervisor route it. Type /direct on the last line of the composer; a picker opens (type to filter), and the specialist you choose handles that message. The pin applies to that one send. (/direct isn't available in Deep Traversal.)
Attachments
Click the paperclip to attach files to a message. Supported types are PDF documents and images (JPEG, PNG, GIF, WebP), within these limits:
| Kind | Max count | Max size each |
|---|---|---|
| Documents | 5 | 4.5 MB |
| Images | 20 | 3.75 MB |
Documents can be attached in both Express and Deep Traversal modes: they are saved to the conversation store, and the coder reads and analyzes them. Images use vision and remain Express-only — if you switch a message that has images attached to Deep Traversal, the images are removed and a note explains why.
Composing and sending
- Press Enter to send; Shift+Enter inserts a new line.
- While a response is generating, the composer is locked until the run finishes.
Runs keep going if you leave
Chat runs — both Express and Deep Traversal — are detached: they execute server-side and keep running even if you close the tab, navigate away, or refresh. You won't lose a run by leaving. When you reopen the conversation, Studio rehydrates the full run — its thinking, Thought Map, workspace files, and registered datasets — so you pick up exactly where it left off, whether it's still in progress or already finished.
Inspecting the configuration
An info ("i") affordance opens the Configuration Inspector. It has two views you can toggle between:
- Agent map — the Agent Map or Swarm Map for the configuration this conversation is pinned to, in read-only form, so you can see exactly which version, Knowledge Graph, and connectors are in play.
- Datasets — the datasets registered across this conversation, which confirms they persist when you leave and return.
If the agent's active configuration changes after your conversation started, Studio shows an informational config drift note describing what changed. It never blocks your messages; if you'd prefer to use the new configuration, you can start a fresh conversation from the prompt.
What an agent shows you
Replies render with full formatting — Markdown, math, and syntax-highlighted code — and can include inline charts and grids built from datasets the agent produced.
Each message can also expand into panels that reveal how the answer was produced:
- Thoughts — the agent's reasoning behind the response.
- Thought Map — for Knowledge Graph retrieval, an interactive map of the path the agent traversed through the graph, with a replay that re-reveals the steps in order.
- Citations — the documents the answer drew on.
- Web Sources — any web pages consulted.
- Workspace — files the agent created during the run.
Thought Map
When an agent retrieves from a Knowledge Graph, the Thought Map visualizes that retrieval as a graph you can rotate and explore, color-coded by level of detail. Because AMPG works by casting a wide, meaning-based net and then narrowing to the passages that matter for your question (see How retrieval works), the map makes that path inspectable. Use replay to step through how the agent moved through the graph to build its answer.
Workspace viewer
When a run produces files, the Workspace panel shows them in a file tree with sizes and download links, including the location of the workspace. While the run is active the view updates live and stays visible briefly after the run completes so you can see the final files. See Code interpreter and datasets.
After a very long Deep Traversal run, the live screen-share view can take up to about a minute to finalize the run's artifacts before it collapses into the workspace icon. This is expected — the files are being written and indexed — so give it a moment after the run reports complete before assuming anything is missing.
Capacity notices
If Amazon Bedrock is temporarily rate-limited mid-conversation, a notice appears inline in the chat explaining the slowdown, separate from the global status banner.
Related pages
- Agents and Swarms — what you chat with.
- Code interpreter and datasets — the workspaces, grids, and charts you see in chat.
- Workflows — for longer, repeatable, multi-step work.