You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
start_agent without detach: true now joins: the child's result lands in
the parent run directly, and finish refuses to end done while a joined
child is still unfinished (suspends instead of asking the model again). detach: true keeps the old always-replies behavior.
call_agent's unified path starts a delegate as a child task and waits on
it inline, suspending instead of blocking once limits.inlineWait passes; via: "a2a" opts a delegate out to keep the plain synchronous A2A call,
which is the only way delegation survives a full Hub outage today.
PauseTask/ResumeTask (agen pause/resume), with tree: true to hold
or release a whole task tree; a held task is skipped by leasing and a
running one suspends at its next turn.
CancelTask(mode: "graceful", grace_period_seconds) gives a run (or a
whole tree) a last turn to call finish before the existing hard cancel
reaches it; CancelTask(tree: true) cancels a whole tree in one call.
on_parent_end: a task's own unfinished joined children are cancelled
when it ends (Temporal-style parent-close), scoped to joined children
only — start_agent's async children are never touched by it.
report_progress, QueryTask (progress, outcome, a simplified activity
cut) and agen query; signal_agent/query_agent for a caller's own
descendants, delivered into a running run via its task inbox before its
next model call.
GetTaskTree/agen tree <task-id>, Task.root_task_id, and a task tree
view in the console's task detail panel.
Live trust now propagates through call_agent's via: "a2a" path too,
closing a gap where a tainting call could be bypassed that way.
Queue modes
conversations.queueMode: followup (default), collect (merges a new
submission into the conversation's still-queued task instead of queuing
separately, with idempotency-aware merging), steer (delivers into the
running task without starting a new one) and interrupt (stops the
running task after a grace period, then queues the new one).
SubmitTask.delivery and event coalescing close out the queue modes'
remaining documented gaps.
Budgets
A tree budget (overrides.budget on the root task) narrows every
descendant's effective budget to the tree's remaining spend, re-checked
live on every turn so a sibling's concurrent spend is never missed.
Per-deployment label budgets (labelBudgets: {"key=value": maxUsdPerDay})
exclude only the candidates carrying an exhausted label from leasing;
everything else in the deployment's queue is unaffected.
budget.threshold events at 50/80/100%, GetMetrics budget rollups, a
model-visible status line past 50% of any limit, cached-input token
weighting, and a deployment-wide hold.
Memory, artifacts and locks
block: memory providers, memory_update (CAS), SetMemoryBlock/ ListMemoryBlocks, and live url: context sources.
save_artifact/read_artifact built-ins plus GetArtifact/ ListArtifacts, so a tool result doesn't have to crowd the context.
Declared locks (SubmitTask.locks, waited on all at once) and mid-run acquire_lock/release_lock with fencing tokens.
Per-task overrides narrow a task's tools, skills, budget and prompt.
Task workspaces (a per-task directory) and permission rule conditions.
Trust, eval and feedback
Trust and taint (operator > trusted > untrusted): trustBelow rules and
live taint propagation from start_agent.
agen replay (opt-in interaction recording, re-run a recorded run
offline) and agen eval (run scenario files against a bundle with
mocked tools, pass@1/pass^k), plus deploy --require-eval and agen eval history.
RecordFeedback/ListFeedback: judgements on tasks, with approvals
counted as feedback too.
Scheduling and triggers
Cron triggers: jitter, catchUpWindow, overlap (skip/buffer_one), minStartSpacing. Random triggers (count, window, minGap, days).
Fallback providers for the model-routing side of the same slice.
notBefore/schedule_key (atomic replace) and task priority (lease
order within a deployment's queue); scale.reserve and stopping
background-priority work first under budget pressure.
Events, notifications and run outcomes
A resumable event log (ListEvents, agen events) and notifications
from it, including budget thresholds, as events.
Run guards: identical-error and stuck-run detection, cut answers, and
hidden denied tools; the finish tool, a result schema, and reporting
at a limit instead of failing outright.
Conversation rollover: a running summary replaces dropped history once
a run outgrows contextTokens, instead of just truncating it.
Protocol and tools
A2A 1.0 and MCP 2025-11-25 version bumps with negotiation.
agen/trust on every MCP tool call's _meta; an idempotent or
read-only MCP tool call is retried once after the server dies instead
of failing the whole run.
Fixed along the way
A data race between the scheduler's fire-and-forget background work
(notification delivery, task-end hooks) and the test clock seam, caught
by go test -race after the hold/join/tree-budget work added two more
per-turn Hub round trips alongside it.
The live tree-budget check's Hub round trip (and the other two sharing
its pattern) shortened from a 10s to a 3s timeout, so an unreachable Hub
costs a turn 3s per check instead of stacking three full 10s waits.