Changelog

What shipped

Every release to the runtime, SDK, and control plane. Breaking changes are called out explicitly and always ship behind a version bump.

Jul 28, 2026

v1.8.0

Added

Multi-agent workflows

Compose agents into a supervisor graph. Child runs inherit the parent trace, so a delegated task is one span tree rather than four disconnected ones.

Graphs are declared in code and versioned alongside your agents. A supervisor can fan out to children in parallel or sequence them, and the retry budget is enforced across the whole tree rather than per node.

Delegation respects the parent’s retry budget and handoff policy. If a child trips a handoff condition, the whole workflow escalates rather than the child failing silently.

Jul 28, 2026

v1.8.0

Added

Multi-agent workflows

Compose agents into a supervisor graph. Child runs inherit the parent trace, so a delegated task is one span tree rather than four disconnected ones.

Graphs are declared in code and versioned alongside your agents. A supervisor can fan out to children in parallel or sequence them, and the retry budget is enforced across the whole tree rather than per node.

Delegation respects the parent’s retry budget and handoff policy. If a child trips a handoff condition, the whole workflow escalates rather than the child failing silently.

Jul 28, 2026

v1.8.0

Added

Multi-agent workflows

Compose agents into a supervisor graph. Child runs inherit the parent trace, so a delegated task is one span tree rather than four disconnected ones.

Graphs are declared in code and versioned alongside your agents. A supervisor can fan out to children in parallel or sequence them, and the retry budget is enforced across the whole tree rather than per node.

Delegation respects the parent’s retry budget and handoff policy. If a child trips a handoff condition, the whole workflow escalates rather than the child failing silently.

Jul 11, 2026

v1.7.2

Fixed

Retry budget no longer double-counts validation failures

A rejected tool call was consuming two retries instead of one. Agents on tight budgets were giving up roughly a third earlier than configured.

The regression landed in v1.7.0 with the stricter schema validator and affected any agent whose tools declared required fields. No migration is needed.

Budgets now decrement once per attempt. No change is needed on your side — existing agents will simply use the budget you configured.

Jul 11, 2026

v1.7.2

Fixed

Retry budget no longer double-counts validation failures

A rejected tool call was consuming two retries instead of one. Agents on tight budgets were giving up roughly a third earlier than configured.

The regression landed in v1.7.0 with the stricter schema validator and affected any agent whose tools declared required fields. No migration is needed.

Budgets now decrement once per attempt. No change is needed on your side — existing agents will simply use the budget you configured.

Jul 11, 2026

v1.7.2

Fixed

Retry budget no longer double-counts validation failures

A rejected tool call was consuming two retries instead of one. Agents on tight budgets were giving up roughly a third earlier than configured.

The regression landed in v1.7.0 with the stricter schema validator and affected any agent whose tools declared required fields. No migration is needed.

Budgets now decrement once per attempt. No change is needed on your side — existing agents will simply use the budget you configured.

Jun 24, 2026

v1.7.0

Added

Eval suites on pull requests

Point the CLI at a suite and Baton runs it against the changed agent on every pull request, diffing scores against the last green build.

Suites live beside your agent definitions and run inside the same job as the rest of your tests, so there is no extra infrastructure to stand up.

Thresholds are declared per case. A regression below the floor fails the job with the offending run linked, so reviewers land on the trace rather than a log dump.

Jun 24, 2026

v1.7.0

Added

Eval suites on pull requests

Point the CLI at a suite and Baton runs it against the changed agent on every pull request, diffing scores against the last green build.

Suites live beside your agent definitions and run inside the same job as the rest of your tests, so there is no extra infrastructure to stand up.

Thresholds are declared per case. A regression below the floor fails the job with the offending run linked, so reviewers land on the trace rather than a log dump.

Jun 24, 2026

v1.7.0

Added

Eval suites on pull requests

Point the CLI at a suite and Baton runs it against the changed agent on every pull request, diffing scores against the last green build.

Suites live beside your agent definitions and run inside the same job as the rest of your tests, so there is no extra infrastructure to stand up.

Thresholds are declared per case. A regression below the floor fails the job with the offending run linked, so reviewers land on the trace rather than a log dump.

Jun 2, 2026

v1.6.1

Security

Scoped API keys and per-tool egress rules

Keys can now be scoped to a single agent or environment, and each tool contract can declare an allowlist of hosts it may reach.

Existing keys keep their current access until you scope them, so nothing breaks on upgrade. Scoping is available from workspace settings and through the management API.

Each tool contract can additionally declare an allowlist of hosts it may reach. Calls to anything outside the list are blocked before the handler executes and surface as a policy span in the trace.

Jun 2, 2026

v1.6.1

Security

Scoped API keys and per-tool egress rules

Keys can now be scoped to a single agent or environment, and each tool contract can declare an allowlist of hosts it may reach.

Existing keys keep their current access until you scope them, so nothing breaks on upgrade. Scoping is available from workspace settings and through the management API.

Each tool contract can additionally declare an allowlist of hosts it may reach. Calls to anything outside the list are blocked before the handler executes and surface as a policy span in the trace.

Jun 2, 2026

v1.6.1

Security

Scoped API keys and per-tool egress rules

Keys can now be scoped to a single agent or environment, and each tool contract can declare an allowlist of hosts it may reach.

Existing keys keep their current access until you scope them, so nothing breaks on upgrade. Scoping is available from workspace settings and through the management API.

Each tool contract can additionally declare an allowlist of hosts it may reach. Calls to anything outside the list are blocked before the handler executes and surface as a policy span in the trace.

May 19, 2026

v1.6.0

Changed

Traces now stream instead of buffering

Spans appear in the trace view as they complete rather than after the run finishes. Long agent runs are debuggable while still in flight.

This changes the SDK’s default flush behaviour. If you were polling for completed runs, you can subscribe to the span stream instead and drop the poll loop.

The export path is unchanged — OpenTelemetry batches still ship on the same interval, so existing collectors need no reconfiguration.

May 19, 2026

v1.6.0

Changed

Traces now stream instead of buffering

Spans appear in the trace view as they complete rather than after the run finishes. Long agent runs are debuggable while still in flight.

This changes the SDK’s default flush behaviour. If you were polling for completed runs, you can subscribe to the span stream instead and drop the poll loop.

The export path is unchanged — OpenTelemetry batches still ship on the same interval, so existing collectors need no reconfiguration.

May 19, 2026

v1.6.0

Changed

Traces now stream instead of buffering

Spans appear in the trace view as they complete rather than after the run finishes. Long agent runs are debuggable while still in flight.

This changes the SDK’s default flush behaviour. If you were polling for completed runs, you can subscribe to the span stream instead and drop the poll loop.

The export path is unchanged — OpenTelemetry batches still ship on the same interval, so existing collectors need no reconfiguration.

Create a free website with Framer, the website builder loved by startups, designers and agencies.