Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
JavaScript agentagentsai-agentdeepagentsdesignhyperframeslanggraphmanaged-agentsmotion-graphicspaperslackvideo
Upstream asks
Seams this project works around today, each with what it does instead. Partner questions for HyperFrames and Paper are in docs/partners/.
Managed Deep Agents (0.9.4)
| Ask | Why | Today |
|---|---|---|
| A public "start a run on this thread now" API | waking the lead after background work lands in 60–120 s through one-time schedules, which round to the next minute | runtime.schedules.create(at=now+61s, new_thread=False) (orchestration) |
| Let a scheduled run end without a post | schedule delivery posts the newest non-empty assistant text of the thread, from a hook installed ahead of every authored middleware; an empty reply would repost an earlier one | a [quiet] reply becomes a one-line status of the running work |
| Async subagents that share the lead's sandbox, and a background general-purpose subagent | helpers must work in the lead's sandbox on the lead's files | in-process helpers behind HelperRunner, with heartbeats and durable transcripts |
| Schedules creatable for a thread from outside a run | after a restart, a thread has no route for wake-ups until a run happens on it | a sentinel schedule per busy thread, pushed forward while the process lives |
Streaming and message updates on runtime.channel |
live progress and the footer need updates to one message | the Slack Web API with the reserved mda/slack-bot-token connection |
| A way to add input to a running run, or choose the strategy for a busy thread | a second Slack message queues behind the running turn | steering reads the thread before each model call and folds follow-ups in; the queued run ends with one line |
| Route direct-message follow-ups to the thread already working | a new top-level DM is its own thread and runs in parallel | not handled |
| Expose a schedule run's Slack conversation and thread | after a restart, wake-up turns cannot find their Slack thread until a message run happens | progress reuses the thread's last known target |
| Dedupe Slack's duplicate deliveries at ingress; react when a message is queued | duplicate events and a silent queue | not handled on MDA (the self-hosted adapter dedupes) |
An ignore list for mda deploy |
it archives Git-ignored files (except .env and virtualenvs), so renders in workspace/out/ count toward the 200 MB limit |
doctor warns past 150 MB; clear old renders |
| Same-thread wake-ups for API callers: agent-owned schedules without a channel, or API keys counted as people | a thread started with an API key cannot be woken (User schedules require an authenticated person; agent-owned same-thread schedules need a channel), so background work over the API never reports back |
Slack only (deploy guide) |
A larger filesystem for docker_image bakes (honor fs_capacity_bytes) |
the computer is 1.3 GiB; docker_image bakes get a fixed 1 GiB, so a fresh clone can't go straight to mda deploy |
a published snapshot (design-agent sandbox publish) once per workspace |
LangChain (1.4.6)
| Ask | Why | Today |
|---|---|---|
jump_to: "end" from an after_agent hook should reach the graph's end |
it routes back to the first after_agent hook, so a hook that jumps loops until the recursion limit |
no after_agent jumps; the quiet reply is rewritten instead |
Deep Agents (0.7.28)
| Ask | Why | Today |
|---|---|---|
| A TTL option for the prompt caching it installs | it always appends Anthropic caching with a 5-minute TTL; a lead read again an hour later pays to rewrite its prefix | CachedChatAnthropic sets every breakpoint to the role's TTL |
| A way to shorten skill descriptions in the resident list | vendored skills carry long upstream descriptions | SkillListMiddleware swaps them per request from config/vendored-skills.toml |