eve

Where it stands, 2026-10-13

Eve 0.79.0 can accept an answer before it has asked the question: queued approve text can settle a later approval prompt. The repair was main-only on 13 October 2026. Keep consequential tools off this path until a release includes the repair and your deployment rejects pre-prompt approval text.

Install npm install eve@0.79.0, released 13 October 2026, with Node 24 or later. Vercel’s TypeScript framework defines agents as reviewable files. This non-prerelease package release remains beta, with APIs and behavior subject to change.

Serve Eve tools over MCP under the caller’s identity.

Opt in with mcpChannel({ tools: true, skills: true }). Eligible calls run as the route-authenticated caller or a trusted forwarded user; conversation-only tools are excluded. Another harness can use Eve’s tools without an Eve model turn. Before replacing a private bridge, spend the review on which callers each tool admits: agent instructions do not govern these calls, and calls needing human approval return approval_required without running.

Parallel calls and lost-response retries do not supply ordered, exactly-once effects. Treat MCP readOnlyHint as a server assertion; enforce access in tools or policies.

Audit approvals that borrowed another person’s account. 0.78.0 changed approved execution to the requester’s identity, connections and credentials, with a policy recheck. For 0.70.0 through 0.77.0, compare approval.settled.responderPrincipalId with the requester and inspect mismatched calls’ effects; through 0.71.2, inspect later calls in that turn too. Redesign workflows that relied on an approver’s credentials: approved calls now use the requester’s identity and credentials. The explicit approver-credential proposal was unmerged on 13 October 2026.

Migrate live sessions and clients together. From 0.63.0, keep older deployments until their sessions end: 0.69.0, 0.70.0 and 0.75.0 changed checkpoints and block backward handoff across their boundaries.

If upgrading from 0.63.0, follow the migration record for configuration, CLI and custom-provider changes before moving clients or extensions.

Replace guards that throw. Since 0.68.0, hook-handler failures, including metrics failures, are logged while work continues; guards that must stop the turn need ctx.cancel(). Inspect entitlement and credential hooks. Test for turn.cancelled.

Choose who may answer and test the whole wait. Approvals hold the same turn. Without an approval.response policy, only an authenticated requester answers; anonymous or unidentified requesters allow any responder. Set a policy for designated approvers. Test that denying approval prevents the tool from starting. Include cancelled responses and rejected child responders and cancellation after resume.

Disable local editing on shared development hosts. eve dev mounts local self-modification by default, and every admitted caller can use enabled editing tools. Use eve dev --no-default-extensions and inspect agent/extensions/: explicit mounts remain enabled. Check the listener: explicit host or Nitro hostname takes precedence over the 127.0.0.1 fallback. For deployed editing, use eve/self-modification/remote, which requires authorization and opens draft pull requests.

Before sensitive evals, choose what leaves your deployment. Configured Datadog reporters export inputs, outputs, expectations, errors and assertion details by default. Set the relevant recordInputs, recordOutputs, recordExpectedOutputs, recordErrors and recordAssertionDetails options to false. Review eval exports separately from runtime audience-based content capture. On new Vercel projects, Eve attempts 100% trace sampling but warns if setup fails; inspect actual settings. Update custom reporters: optional passed means failure only when passed === false.

Review source changes before resuming local work. Restarts leave earlier workflows dormant; eve dev --resume may perform work before replay fails after source changes. Choose recovery deliberately. Do not assume durability means automatic restart recovery.

Test the smaller default tool set before retiring planning code. Eve removed todo and made ask_question opt-in; catalog tools appear on demand, then stay. Compare accepted outcomes and human review. Keep planning code where you need the oversight. Deletion alone does not measure model progress.

Try the coding extension in Eve before considering a harness switch. Mount eve/extensions/code for repository work. The vendor summary generated on 3 October 2026 reports 24/36 resolved attempts for Eve source 4d1491477a2e4f220131a3eb857a030c9eec0419, measured on 30 September: 12 tasks, three attempts each, on anthropic/claude-sonnet-5.5. The same vendor benchmark reports Pi 0.85.1 at 26/36 and opencode 1.18.30 at 18/36, alongside Eve’s 24/36; their reported 95% intervals overlap. By the vendor’s stated method, that establishes no winner.

Overlap does not establish equivalence. The tested Eve source is not 0.79.0. These numbers do not justify a harness switch.

What is unresolved

• Keep automatic approval to reversible effects. auto() is opt-in and sends tool input to its decision-model provider. The research note does not establish how often automatic approval mistakenly permits a tool call that should be denied. Check that tool calls your policy must forbid never execute in your deployment. Use deterministic entitlement checks for irreversible effects.
• Wait for a release fixing result matching between connections on the same host. The MCP/OpenAPI toolResultFrom repair is main-only. A release with the repair and hook/result tests would settle matching for your distinct connections. Identical URL and description leave fallback matching ambiguous; the original definition objects loaded by Eve still distinguish the connections; test those separately even after the release.
• Require maintained source before choosing Buzz ACP. The adapter source was removed while the 0.0.46 npm artifact survives. That artifact does not prove ongoing maintenance. A maintained source path and supported release with Eve integration tests would settle the choice.
• Prove recovery before choosing a non-Vercel host. The hosting guide requires an explicit workspace-peer transport for delegation outside Vercel and still names the beta Workflow line, while Eve’s Workflow build inputs use stable 5.x pins. Neither establishes cross-host recovery. Before choosing another host, require a supported Workflow backend configuration using those stable 5.x dependencies and passing crash and replay tests.

Evidence and limits

MCP tools and skills are released but opt-in; automatic approval is released but opt-in; experimental schedules are released but opt-in.

This 2026-10-13 profile draws on source, documentation and attributed vendor measurements. It asserts no end-to-end deployment or recovery result. See METHOD.md for the evidence contract.