Changelog
What shipped, and when.
Every release, fix and security change, with the version number you’ll see in your lockfile.
4.2.0
Feature
v4.2 · Replay from any step, in bulk
Select hundreds of failed runs and replay them from the step that broke, with the fix you just deployed.
4.1.0
Feature
v4.1 · Approvals in Teams and email
Human-in-the-loop requests can now be answered from Microsoft Teams and email as well as Slack.
4.0.3
Improvement
v4.0.3 · Faster cold starts in eu-west
Median step start latency in eu-west dropped from 180 ms to 41 ms.
4.0.0
Feature
v4.0 · Fair queueing per key
Concurrency limits can now take a key, so one tenant’s burst can’t starve the rest.
3.9.2
Fix
v3.9.2 · Fix: sleep durations over 30 days
Sleeps longer than 30 days could wake up to an hour early. They now wake on time.
3.9.0
Feature
v3.9 · Token and cost tracking per step
Every LLM step now records input and output tokens and cost, with budgets and alerts per workflow.
3.8.0
Feature
v3.8 · Python SDK 1.0
The Python SDK is stable, with async support and typed events.
3.7.4
Security
v3.7.4 · Security: scoped API keys
API keys can now be limited to specific environments and actions.
3.7.0
Feature
v3.7 · MCP tool servers
Agents can call tools from any MCP server, with each call recorded as a step.
3.6.1
Improvement
v3.6.1 · Dashboard: trace search
Search every step’s input and output across runs, in seconds.
Roadmap
What we’re building next.
Dates are targets, not promises. Vote on the public board and we’ll tell you when it ships.
Now · Q4 2026
Workflow versioning in the dashboard
See which code version each run started on and pin long runs to it.
Next · Q1 2027
Rust SDK
The fourth SDK, with the same step API and one shared runtime.
Later · 2027
On-device runs
Run lightweight workflows at the edge and sync traces home.
Updates
Get the release notes by email.
One short email when something ships. No marketing, and you can reply to the engineer who wrote it.