Roadmap
New work should make runs easier to control, inspect, and reproduce — not hide execution behind magic.
Runtime#
- Richer retry policy primitives with safe defaults for reads vs writes
- First-class timeout configuration at provider and tool edges
- Broader approval adapters (remote reviewers, queues)
- Clearer output repair controls after validation failure
Studio#
- Trace comparison across runs
- Saved views / filters for operators
- Approval audit history
- Memory diffs between turns
- Tool latency breakdowns tied to the timeline
Providers#
- Additional vendor adapters
- Capability metadata (tools, vision, structured output) for routing
- Normalized usage reporting across adapters
Observability#
- OpenTelemetry exporter
- Stable JSON trace snapshots for CI fixtures
- External sinks without forking core
Principles#
Prefer features that expose structure over features that conceal it. If a change cannot be explained in a trace or shown in Studio, revisit the design.