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
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Models and providers
harness.json: fallbackModels (tried in order when a model keeps failing
or gives no answer), reasoningEffort, openrouter (provider routing
and plugins, sent as given) and router for routing models
(typesafe/jev-router with tiers low to high, openrouter/auto and openrouter/auto-beta with tiers low to max): tier, maxTier, budgetTier, allowedModels, excludedModels. Tiers are checked per
router at deploy and start.
Model-call spans record the model that served the call
(gen_ai.response.model), the router's tier and decision, and fallbacks.
An empty answer is asked again once, then falls back; a run that gets no
answer from any model fails with no_answer.
Account errors (401, 402, 403; not a 403 for input flagged by moderation)
are never retried or failed over. The Hub
holds the provider account: its deployments take no new tasks, do not
scale up or wake, and one probe goes through after a backoff (5 min,
doubling to 1 h); a run that succeeds ends the hold. ListProviderHolds, ResumeProvider, agen providers [resume], and provider.held / provider.resumed notifications.
Every way to add an agent can pick a routing model: agen init --model --tier --max-tier --fallback, the console's New agent and Templates
forms, and the Python, Node and Go SDKs (router, fallback_models, reasoning_effort, openrouter, context_tokens; a run's model_tier), covered by the shared conformance cases.
The Python, Node and Go SDKs take maxTokensPerRun (the core already did).
SDK agents refuse misspelled settings and run tiers, and inline settings
next to a bundle (the bundle sets them), instead of ignoring them. Node's
types include modelTier; Go keeps tools in the order added, refuses a
duplicate name and returns a closed error after Close; Python's arun
takes the same options as run.
Tasks take a model tier (agen run --tier, modelTier, A2A metadata agen.model_tier) and a start time (agen run --at / --in, notBefore); a task waiting for its start time holds up nothing.
At 80% of the daily budget, leased tasks carry budget_pressure and
routing models use the bundle's budgetTier.
agen wait <task>; agen run --timeout 0 waits without a limit.
A conversation key read back from a task can be sent again as is.
Cron triggers carry nextDueAt and lastDueAt; ListTasks with scheduled; ListTriggerEvents across deployments and namespaces.
Delegation
start_agent: delegate without waiting. The task is queued on the
delegate and the result comes back as a task in the caller's conversation
(source reply:<agent>). A run started over A2A can start child tasks too.
A2A calls wait up to 10 min for a free instance (asleep, starting, busy or
waiting for Nest capacity) instead of failing after 60 s, and a running
call keeps its deployment awake. call_agent waits up to 75 min.
A run on an instance that is draining after a scale-down or redeploy can
keep delegating.
Idle deployments give their capacity back after 1 min while another
deployment waits for room on a full Nest.
Deleting a deployment ends its unfinished runs.
Engine
contextTokens (default 64000): a long run's oldest tool results are
shortened in what is sent to the model once the run outgrows it; the Store
keeps them whole.
agent.mdpreloadSkills: skills whose instructions are always in the
system prompt.
Failed tool spans carry error.message.
Console
Overview: a compact Needs attention (counts by kind, the full list in the
Inbox), a live fleet map (entry points, calls, async starts and results,
self-calls, travelling dots; pan, zoom and find for large fleets), Now
(every run in progress and what it is doing), and a per-agent timeline.
Schedules: cron triggers with their next and last run and Run now, tasks
waiting for their start time, webhooks with New secret, and missed and
refused events.
Costs by served model, and held provider accounts with Resume.
Deployments: bundle versions with roll back; tasks with start time, model
tier, where an async result goes, and Cancel; the run form takes a start
time and a tier. Fleet: add a Nest (join command) and revoke one.
API
GET /api/v1/events sends step events (a model or tool call starting or
ending).
GetActivity: unfinished runs and what they are doing, queues, and
warnings (near limits, quiet runs, long approval waits, queues waiting or
blocked by an approval, bursts, model errors, held providers, missed
schedules, refused webhooks).
GetMetrics: per served model, entry points, and async calls, calls in
flight and p95 per edge.
Tools and templates
Built-in tool servers: agen mcp workspace --root DIR [--no-run] (read,
list, grep, edit, write and run, confined to one folder) and agen mcp sql [--dsn-env NAME] (read-only SQL on Postgres or SQLite).
Instances get AGEN_BIN, so "command": "agen" works without agen on PATH.
New templates: coder, ci-fixer, support, incident-responder and
data-analyst. fleet-steward also reports live warnings and writes the
daily cost report. cost-watchdog and support-triage are removed.
Testing
spec/features.yaml lists every feature and the surfaces it must work on
(SDKs, API, MCP, CLI, console, agent-to-agent calls, installers, Docker,
Kubernetes). Tests tag what they cover, and a check fails when an API
method or field, bundle key, CLI command or SDK option belongs to no
feature, or a feature is untested on a surface without a recorded gap. The support matrix is generated from it.
Scenarios (spec/scenarios) are written once as API calls and played
through the API, MCP and the CLI.
agen notify set --event lists every event, provider events included.