Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
Surfaces: Slack, the API and the CLI
Every surface reaches the same agent. What differs is presentation: in Slack the agent shows its work as it goes, takes follow-ups mid-job, and ends each reply with a short footer.
Slack presentation
Code: src/design_agent/surfaces/slack/ (Web API client, who a run answers, step phrases, the
stream) and src/design_agent/middleware/{progress,footer,steering,tracing}.py.
- Live progress. One message per turn shows a card per step in plain studio language
("Studying the brand", "Rendering the teaser"), through Slack's streaming API
(
chat.startStream/appendStream/stopStream), or one message redrawn withchat.updatewhen streaming is refused. It settles to "Done", or is deleted when the turn's reply is[quiet]. Cards never carry argument text: no paths, queries, file ids or anybody's words. Each background helper streams its own message. - Footer. Only what changes: agents still running, time so far, estimated cost of the job, and
a trace link when
DESIGN_AGENT_TRACE_HOSTis set (a trace URL exposes workspace and run ids, so it is off by default). - Steering. A follow-up while helpers run starts a turn at once (the lead ends its turn after
handing work off), sees what is running, and can pass the note to a helper with
steer_task; the helper reads it before its next model call. While the lead's own turn is busy, it reads new messages in the thread before each model call and folds them in; "stop", "cancel" or "never mind" ends the turn at its next step. - Trace tags. Each run is tagged with its trigger, surface and the kinds of files delivered, and
names the template and version it came from (
DESIGN_AGENT_TEMPLATE_TELEMETRY=0turns that off); no user ids by default.
The bot token stays on the host: progress and footer call the Slack Web API from middleware, and the sandbox never sees a token, a Slack URL or a Slack payload.
Managed Deep Agents
MDA's channel posts the final reply, reacts to the starting message, transfers attachments to
/workspace/attachments/ and provides upload_files. Presentation uses the reserved
mda/slack-bot-token connection and the event on runtime.channel.
A second message in a busy Slack thread becomes a run queued behind the current one (there is no public way to add input to a running run). The steering middleware folds it into the running turn by reading the thread, so the queued run ends at once with "Folded that into the work above." (it has to say something: MDA posts a scheduled or queued run's newest assistant text). A new top-level direct message is its own thread and runs in parallel; only the platform can route it to the working thread. Upstream asks: docs/upstream.md.
Self-hosted Slack
design-agent slack runs a Bolt app in one process with the agent (Socket Mode by default, or
signed HTTP on /slack/events): mentions and direct messages start turns, duplicate deliveries
start one, attachments are fetched on the host with the bot token, finals upload to the thread,
and wake-up replies post there. The same middleware draws progress and the footer. Manifest:
config/slack-manifest.example.yaml.
The API and the CLI
The API returns each turn's reply and delivered files, and background news through
GET /threads/{id}/updates (self-hosting). The CLI prints each step's tool
name as it runs and stays open while background work continues.