February 2026 Changelog
Here is everything that shipped in February.
Scoped Memory (Memory v2)
Agents now support three memory scopes: user, session, and agent. Context persists automatically across runs with zero configuration. Vector search queries all three scopes simultaneously.
Read the full announcement: Introducing Memory v2.
Parallel Tool Execution
The runtime now detects independent tool calls within a single step and executes them in parallel. Average speedup on fan-out workflows is 2.8x. No configuration required — the runtime infers independence from the execution graph.
New Trace Timeline UI
The trace viewer has been redesigned. The new timeline view shows steps horizontally with parallel branches visible as stacked lanes. Each step displays model, tokens, cost, and latency inline. Click any step to expand the full prompt and response.
8 New Built-in Tools
Linear — create and update issues, query projects
Notion — read and write pages, query databases
Jira — create tickets, update status, query boards
HubSpot — create and update contacts, log activities
Twilio — send SMS, make calls, check delivery status
SendGrid — send emails, manage templates
Snowflake — query tables, describe schemas
BigQuery — run queries, export results
All tools include typed inputs and outputs with auto-generated schemas.
Runtime Improvements
Retry backoff is now configurable per tool (previously global only)
Agent runs now support a
timeoutfield at the run level (previously step-level only)The
interruptOnconfig now accepts regex patterns for flexible matching
Bug Fixes
Fixed a race condition in parallel tool execution that could cause duplicate tool calls under high concurrency
Fixed memory TTL not being respected when using the Redis backend
Fixed trace export missing the final step when the agent run was interrupted
Fixed cost attribution double-counting retried model calls
Breaking Changes
None.
Upgrade
All changes are backward-compatible. Memory v1 data is migrated automatically on upgrade.
