Agen
Reference / Support matrix

Reference

Support matrix

What works where, and what proves it. ✓ is tested on that surface, ◐ is offered there but not tested yet (or, with a note, not fully offered yet), – is deliberately not offered there (see the note), and an empty cell does not apply.

MCP tools are generated from the same API methods, so a feature on the API is reachable over MCP; the MCP column shows where that is also tested. Platform tests run on Linux, macOS and Windows in CI.

Core (embed an agent in your program)

Feature Rust Python Node Go
Instructions, model and provider ✓ ✓ ✓ ✓
Agents from bundles ✓ ✓ ✓ ✓
Skills (on demand and preloaded) ✓ ✓ ✓ ✓
Fallback models ✓ ✓ ✓ ✓
Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) ◐1 –2 –3 –4
Routing models (Jev Router, Auto Router) and cost tiers ✓ ✓ ✓ ✓
Reasoning effort ◐5 ◐6 ◐7 ◐8
OpenRouter request settings (provider routing, plugins) ✓ ✓ ✓ ✓
Long-run compaction (contextTokens) ◐9 ◐10 ◐11 ◐12
Limits (turns, tokens, durations, queue) ✓ ✓ ✓ ✓
Run outcomes (finish, result schema, report at a limit) ✓ ✓ ✓ ✓
Host tools (functions in your program) ✓ ✓ ✓ ✓
Permissions and approvals ✓ ✓ ✓ ✓
Run tasks and wait for results ✓ ✓ ✓ ✓
Cancel a task or run ✓ ✓ ✓ ✓
Stream output while a run is in progress ✓ ✓ ✓ ✓
Conversations and labels ✓ ✓ ✓ ✓
Per-task cost tier ✓ ✓ ✓ ✓
Runs, traces and transcripts ✓ ✓ ✓ ✓

Fleet (run agents on the Hub and Nests)

Feature API MCP CLI Console Agent to agent
Instructions, model and provider ✓ ✓ ✓ ✓
Agents from bundles ✓ ✓ ✓ ✓
Skills (on demand and preloaded) ✓ ✓ ✓
Fallback models ✓ ✓ ✓ ✓
Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) –13 –14 –15 ◐16
Routing models (Jev Router, Auto Router) and cost tiers ✓ ✓ ✓ ✓
Reasoning effort ✓ ✓ ✓
OpenRouter request settings (provider routing, plugins) ✓ ✓ ✓
Provider account holds and resume ✓ ✓ ✓ ◐17
Long-run compaction (contextTokens) ✓ ✓ ✓
Limits (turns, tokens, durations, queue) ✓ ✓
Run outcomes (finish, result schema, report at a limit) ✓ ✓ ✓ ✓ ✓
Event log (ListEvents, event stream with a cursor) ✓ ✓ ✓ –18
Host tools (functions in your program) –19
Tools from MCP servers in the bundle ✓ ✓
Built-in workspace and SQL tool servers ✓
Permissions and approvals ✓ ✓ ✓ ✓ ✓
Record and list feedback on tasks (a person, a check, or a judge model) ✓ ✓ ✓ –20
Memory blocks (durable, named text an agent reads and updates across runs), live context sources and conversation rollover ✓ ✓ ✓ –21
Artifacts (save_artifact / read_artifact; a long result need not stay in context) ✓ ✓ ✓ –22
Run tasks and wait for results ✓ ✓ ✓ ✓
Every task of the tree a task belongs to (its root and descendants) ✓ ✓ ✓
How a busy conversation’s queue behaves (followup, the default, queues separately; collect merges; steer delivers into the running turn; interrupt gracefully cancels it first); SubmitTask.delivery overrides it for one submission ✓ ✓ ✓
An agent reports its own progress; query a task’s progress, activity and outcome without waiting on it ✓ ✓ ✓
Cancel a task or run ✓ ✓ ✓ ✓
Cancel a task gracefully, giving each run a last turn to call finish before it is stopped ✓ ✓ ✓
Hold a task, and resume it; a queued one is skipped by leasing, a running one suspends at its next model call ✓ ✓ ✓
Stream output while a run is in progress –23
Conversations and labels ✓ ✓ ✓
Per-task cost tier ✓ ✓ ✓ ✓
Tasks with a start time, an expiry, or at most one queued per key ✓ ✓ ✓ ✓
Cron, webhook, event and random triggers; a burst of events a trigger matches within coalesce merges into one task ✓ ✓ ✓ ✓
Agents calling agents (call_agent) ✓ ✓ ✓
Start an agent without waiting (start_agent) ✓ ✓
Deploy, update, scale, pause and delete ✓ ✓ ✓ ✓
Check a bundle before deploying ✓ ✓ ✓ ✓
Bundle versions and roll back ✓ ✓ ✓
Start from a template ✓ ✓ ✓ ✓
Daily budgets ✓ ✓ ✓ ✓
Platform secrets ✓ ✓ ✓ ✓
Runs, traces and transcripts ✓ ✓ ✓ ✓
Deployment logs ✓ ✓ ✓ ✓
Record a deployment’s model calls and replay a run from its cassette offline ✓ ✓ ✓
Run scenario files against a bundle with mocked tools; report pass@1/pass^k; eval history per deployment and deploy –require-eval ✓ ✓ ✓ –24
Metrics, costs and live activity ✓ ✓ ✓
Notification targets ✓ ✓ ✓ ✓
API tokens, scopes and namespaces ✓ ✓ ✓ ✓
Local fleet (agen up) ✓
Nests (machines that run agents) ✓ ✓ ✓ ✓

Install and run the fleet

Feature Installer (Linux, macOS) Installer (Windows) Docker cluster Kubernetes
Agents calling agents (call_agent) ✓
Deploy, update, scale, pause and delete ✓ ✓
Local fleet (agen up) ✓ ✓
Nests (machines that run agents) ✓ ✓

Footnotes

  1. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on Rust: tested at the engine level (engine/crates/agen-engine/tests/model_routing.rs), not yet exposed through agen-sdk-core’s own AgentSpec ↩

  2. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on Python: bundle-config only, no SDK surface ↩

  3. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on Node: bundle-config only, no SDK surface ↩

  4. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on Go: bundle-config only, no SDK surface ↩

  5. Reasoning effort on Rust: engine tests only: the scripted test provider ignores reasoning effort ↩

  6. Reasoning effort on Python: engine tests only ↩

  7. Reasoning effort on Node: engine tests only ↩

  8. Reasoning effort on Go: engine tests only ↩

  9. Long-run compaction (contextTokens) on Rust: engine tests only: no SDK case runs long enough to compact ↩

  10. Long-run compaction (contextTokens) on Python: engine tests only ↩

  11. Long-run compaction (contextTokens) on Node: engine tests only ↩

  12. Long-run compaction (contextTokens) on Go: engine tests only ↩

  13. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on API: bundle-config only, no SDK surface ↩

  14. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on MCP: bundle-config only, no SDK surface ↩

  15. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on CLI: bundle-config only, no SDK surface ↩

  16. Fallback providers, each a full account, tried in order only after an account error on the primary (not a transient one, and not just a failing model) on Console: no dedicated UI yet; the generic trace/span inspector already shows gen_ai.provider.name and agen.fallback_provider when one fires, not verified by an e2e test ↩

  17. Provider account holds and resume on Console: a held account cannot be made from the console test; Hub tests cover holds ↩

  18. Event log (ListEvents, event stream with a cursor) on Console: the console follows changes live through the same stream; it has no event log view yet ↩

  19. Host tools (functions in your program) on API: deployed agents get tools from MCP servers in the bundle (tools.mcp) ↩

  20. Record and list feedback on tasks (a person, a check, or a judge model) on Console: no UI yet; use agen feedback or the API ↩

  21. Memory blocks (durable, named text an agent reads and updates across runs), live context sources and conversation rollover on Console: no UI yet; use agen memory or the API ↩

  22. Artifacts (save_artifact / read_artifact; a long result need not stay in context) on Console: no UI yet; use agen artifacts or the API ↩

  23. Stream output while a run is in progress on API: tasks return the final result; follow live steps on /api/v1/events ↩

  24. Run scenario files against a bundle with mocked tools; report pass@1/pass^k; eval history per deployment and deploy –require-eval on Console: no UI yet; use agen eval/deploy –require-eval or the API ↩