Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

eve changelog: what's new each month

Every stable eve release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 2 months covered; the current month updates daily.

October 2026 · so far

15 releases: 0.69.0 → 0.72.1

October brought major workflow improvements to eve, with subagent calls becoming resumable tasks, connection tools no longer breaking prompt cache, and better handling of approvals and sign-ins. The month also added file support across more channels and switched to the AI SDK's new decision-making API.

Action needed
  • Chat SDK bridge's send API changed: replace send({ message, ... }) with send(message, options) and use new respond(inputResponses, { thread }) for input requests.
  • session.streamSubagent(started) is now session.agent(started).stream().
  • evaluate from eve/ai is now decide; auto and t.judge() use ExperimentalDecisionModel instead of evaluation.
  • Extension configuration and durable state now belong to each logical mount; duplicate mounts use independent config and state.

Highlights

  • Subagent calls are now resumable tasks that return a receipt, with replies arriving as task results instead of blocking.
  • Connection tools no longer break the prompt cache by using new connectionsearch and connectionexecute tools instead of modifying the prompt.
  • Approvals and sign-ins now hold turns open like questions, emitting turn.waiting and allowing partial answers instead of completing immediately.
  • Files sent over Chat SDK adapters like Slack, Google Chat, WhatsApp, Teams, and Discord now reach the agent instead of remaining as unreachable URLs.
  • The AI SDK's evaluation API was renamed to decisions; evaluate from eve/ai is now decide.

New

  • ctx.agent(name) in workflow tools returns a session to send messages to named agents with outputSchema and signal options.
  • session.agent(started).stream() replaces streamSubagent for resumable subagent interactions.
  • Bash tool no longer blocks on slow commands: commands running over 30 seconds return status running with their pid.
  • session.compact() method for eval sessions to compact between turns without raw fetch calls.
  • Browserbase web search integration through AI Gateway.
  • Numbered options for askquestion and tool approval prompts on Chat SDK adapters without card support like Photon iMessage.
  • Staging file content validation before each model call to detect overwrites in the sandbox.
  • MCP connections now sign in only when the server requests it, not when a connection has auth but no token.

Improved

  • Turn waiting behavior: turn.waiting now emits whenever eve holds a turn, whether for tasks, approvals, sign-ins, or questions.
  • Approval response policies now run for Cancel as well as Approve, preventing rejectable responders from cancelling others' requests.
  • First local subagent turn now uses the caller's trace; remote turns remain separate roots with dispatch links.
  • TUI log modes now use none, error, warn, debug, and all, replacing stderr and sandbox filters.
  • Frontend hooks and EveAgentStore share one conversation client returning ConversationState with messages, turns, requests, tasks, and sessions.
  • Long-lived sessions move to newer deployments after upgrades, handling handoff checkpoints from eve 0.66.0 onward.
  • Token estimation for compaction now uses UTF-8 bytes instead of string length, correctly weighing non-Latin text.
  • Telegram, Discord, and Teams now remove question and approval buttons once answered and show the outcome.

Fixed

  • Model steps ending early at token limits no longer leave tool calls unanswered.
  • Sessions from eve 0.68 and earlier now hand off to newer deployments even when state keys were removed.
  • Long-running sessions no longer accumulate ever-growing workflow logs; idle sessions move to fresh runs after compaction.
  • Partial answers to some pending ctx.ask() or subagent questions now park the turn with turn.waiting again.
  • Steering messages now end interrupted steps with step.completed before the next step starts.
  • useEveAgent no longer aborts in-flight turns during React Fast Refresh or Strict Mode replay.

September 2026

62 releases: 0.0.6 → 0.0.39

September brought major architectural changes to eve, including workflow redesigns, multi-agent project support, and significant UI refinements. The month focused on improving agent composition, deployment flexibility, and development experience.

Action needed
  • ctx.getSkill() and SkillHandle/SkillFile types removed; use model's sandbox tools for skill files instead
  • Background execution removed from defineTool and dynamic tools; use defineWorkflowTool for durable background tasks
  • agent/instrumentation.ts must be split into path-named files under agent/instrumentation/ directory
  • Deprecated instrumentation compatibility shapes removed; use tracePolicy and object-form export policies

Highlights

  • Multi-agent project discovery and support with agents/ directory structure and per-agent CLI commands
  • Workflow architecture overhaul: sessions now run directly in owning workflow instead of dispatching child turns per message
  • New built-in extensions including computer-use and code capabilities available as modular imports
  • Simplified agent delegation with ctx.agent(target, input) and new workspace agent routing patterns
  • Redesigned development terminal UI with model selection, clearer metadata, and inline catalog completions

New

  • Multi-agent workspace discovery from agents/ directory with agent-scoped CLI commands
  • Computer-use extension as eve/computer-use for agents needing desktop interaction
  • Code extension (eve/extensions/code) bundling applypatch, gh, grep, and coding skills
  • Workflow factory from eve/tools/workflow for running model-supplied JavaScript through durable agent calls
  • taskDeliveryPolicy option (auto or cohort) for controlling when related work reports are sent
  • defineWorkspaceAgent() for delegating to workspace peers by name with automatic routing
  • withEve() from eve/next for composing native workspace agents with authored services
  • Microsoft Teams channel setup with guided flow through eve add channel/teams

Improved

  • Sessions now park after each turn instead of ending, with resumable background work support
  • Dynamic tool callbacks remain isolated by session, lifecycle scope, and resolver during replay
  • Tool approval prompts stay independent; approving one call no longer auto-authorizes other pending calls
  • Message and reasoning appends stored as plain deltas instead of cumulative text for efficiency
  • Background subagents inherit authentication context from the turn that created them
  • Model selection in dev TUI now grouped into speed and reasoning choices with clear defaults
  • Slack responses longer than native limit now uploaded as Markdown snippets
  • Activity observation properly forwarded when background tasks start local or remote subagents

Fixed

  • Approved tool calls silently not executing when task state was injected on resume step
  • Approved file-memory setup failures with Vercel CLI 57 by connecting store without unsupported output flags
  • Sandbox base image bash exit issue causing TERM environment variable problems
  • Vercel Workflow race where rejected inline-step preclaim could skip owner's body and exhaust retries
  • Context key errors when parent has dynamic skills and attempts subagent delegation
  • MCP tool failures with "Cannot read properties of undefined (reading 'push')" when app uses different Zod version

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. eve: pricing, features and alternatives · All changelogs