|
|
|
DISPATCH // 05.04.26
|
|
|
TRANSMISSION
A rebuilt Twill engine.
Hi there, we spent the last few weeks
rebuilding the runtime that drives every Twill task. New:
reasoning level control, token-level streaming, message
editing, and live preview surfaces. Faster: across the board.
|
|
|
Reasoning level control
Pick how much the model thinks before acting. Use a
lower setting for fast mechanical tasks; turn it up when
you want the agent to plan, weigh tradeoffs, and avoid
premature edits.
|
|
Token-level streaming
Agent responses stream in token by token. You see the
model write as it works, instead of waiting for each
turn to land.
|
|
Message editing
Edit any message in a conversation to rewind the run to
that point and continue from there. When a prompt
didn’t land, adjust it in place instead of
starting over.
|
|
Live preview surfaces
Fullscreen preview URL for the running app, a
browser-based web terminal into the sandbox, and live
entrypoint logs so you can watch
pnpm install finish
in real time.
|
|
|
|
|
Fewer network round-trips per turn between the app and the
sandbox. Inside each sandbox, the agent CLI process now stays
warm across turns instead of being relaunched on every
message.
|
|
Sandbox setup
/
~5×
faster
|
|
|
Follow-up response start
/
~4×
faster
|
|
|
Stopping a run, queuing a follow-up, sending a correction. All
of it feels closer to running the CLI locally than to talking
to a remote service.
|
|
|
|
Most of the runtime now lives in
agentbox, an MIT-licensed TypeScript SDK we extracted. It provides
sandbox primitives: command execution, file IO, port
forwarding, repo snapshots, provider adapters. We run on
Modal.
|
|
|
|
|
|
© 2026 Anybase Inc.
|
|
1411 Broadway, 16th Floor, New York, NY
|
|
|
|
