Mastra

AdaptableLogger contract standardizes trace correlation by injecting trace_id/span_id into native log records during traced operations, with loggerOptions controlling correlation and export. PinoLogger implements the adapter. Deployment worker runtime now exposes a /health endpoint (503 during startup, 200 once ready). Also fixed worker schedule discovery, resume robustness with falsy payloads, and broken click-through from Studio. Removed several @mastra/playground-ui DataList API options.

Highlights

New storage backends: Elasticsearch + Valkey (GLIDE)

@mastra/elasticsearch adds an ElasticSearchStore that can power memory, workflow snapshots, scores, and semantic recall from a single Elasticsearch cluster, while new @mastra/valkey and…

New dataset experiment APIs let external orchestrators own the experiment loop while Mastra stays the system of record, with server-side run, external result submission, and idempotent finalization. Concurrent workflow resume() calls are now atomic — losers get WORKFLOW_RESUME_ALREADY_CLAIMED and HTTP 409. Also added graceful shutdown controls for generated servers, automatic while-active message delivery in sessions, and a multi-turn LLM judge scorer.

Highlights

Durable execution for Agents API (no deploy required)

Stored agents created via the Agents API can now run with durable: true, enabling durable execution (with configurable loop settings) without shipping new code, and inheriting cache/pubsub from your…

CostGuardProcessor is renamed to TokenCostControl with new budgeting options (warnAtPercent, per-request maxCost, user/organization/session scopes, per-model breakdowns). Automatic agent runs are now opt-in per Factory on upgrade, and external-parent traces now appear as trace roots in Studio. Also adds Agent.listActiveThreadRuns() and indexed redaction tokens.

Highlights

File-based Agent Schedules

File-based agents can now declare recurring tasks via a schedules/ directory (TS or Markdown) that syncs to schedule storage at startup, eliminating runtime registration code and making schedule IDs stable across builds.

###…

processToolResult lets processors scan, transform, or abort tool output before the LLM sees it, and transient: true signals deliver per-turn context without persisting to stored memory. Also added tool definitions to MODEL_GENERATION spans and the @mastra/deepeval exporter; fixed a suspended-runs memory leak and a session resumption bug.

Highlights

Persistable Declarative Workflows (JSON graphs + rehydration)

Workflows can now be authored as data and round-trip through JSON via new declarative step entry types (agent, tool, mapping, plus composite entries), with toStorableGraph() and…

Code Mode transports can now run without a workspace sandbox via requiresSandbox: false, with a new @mastra/isolated-vm package providing V8-isolated execution. A built-in webSearchTool resolves to provider-native web search for supported models. The ChannelHandler 4th parameter (ctx) is now required — code that calls a ChannelHandler-typed function with only three arguments must be updated to pass ctx.

Highlights

AgentController Chat Channels (Slack/Discord/Telegram, etc.)

AgentController can now run durable agent sessions inside chat threads via Chat SDK adapters, with native streaming output, tool approval cards, typing indicators, and controller-owned webhook…

Channel reaction tools (add_reaction, remove_reaction) are no longer auto-injected into channel-bearing agents — opt in via channels.getTools(). New read-only options for local sandboxes and filesystem restrict the working directory. Fixed sub-agent tool approvals resuming the wrong run, thread title generation receiving messages twice, and isTaskComplete leaking scorer reports into final answers.

Highlights

Sandbox Deployments + One-Step mastra build (New @mastra/deployer-sandbox)

You can now deploy a full Mastra server (including Studio) into networking-capable workspace sandboxes (Vercel/E2B/Daytona) and get a live URL in seconds via the new…

Highlights

Durable Agent Crash Recovery (server + client + boot-time)

Durable agents can now discover and recover orphaned RUNNING runs after a restart via DurableAgent.listActiveRuns(), recover(), and recoverActiveRuns(), plus…

The heartbeats API is renamed to schedules across methods, routes, and types; agent schedule IDs now use the agent_ prefix and the default fire signal tag is <schedule>. New @mastra/livekit package enables realtime voice agents with WebRTC, VAD, STT/TTS, turn detection, and barge-in handled by LiveKit while Mastra drives replies, tools, and memory. File-system routing now auto-discovers storage.ts, observability.ts, server.ts, workflows, and agent processors from src/mastra, and @mastra/deployer can auto-construct a Mastra instance when src/mastra/index.ts is absent.

Storage retention policies can now be declared per table with a safe storage.prune() method to delete aged data across growth tables. Multi-tenant isolation is now available end-to-end for datasets, experiments, scores, and scorers via optional organizationId/projectId scoping at the storage layer. New scoreTrace() and scoreTraceBatch() APIs let you score already-stored traces without re-running agents, with optional provenance metadata linking scores back to dataset items.

Highlights

Multi-session Harness architecture (Session-first APIs)

Harness is now a pure factory/shared-resource owner: create isolated sessions via harness.createSession() (get-or-create by resourceId), with run control, state, event bus, model/mode switching,…

Fixed a WorkOS OAuth login failure caused by a missing PKCE verifier cookie. Agent signal drains now record pending signals through the canonical transcript path and rotate response message IDs consistently, preventing follow-up signal turns from attaching to the wrong assistant response. Also includes faster long-thread resume via optimized state signal restoration and redesigned Studio UI panels.

Highlights

Trusted “system actor” execution for workflows, tools, memory, and agents

You can now run server-side/background work (cron jobs, schedulers, queues) as a trusted actor across workflow.execute(), tool.execute(), agent.generate()/stream(), and memory…