AI agents that keep working when no one is watching.
The Harness turns an agent into a dependable service. It wakes on schedules and events, keeps its state, recovers from crashes and stays inside the limits you set.
An agent that runs on its own.
Wake at six, check the pipelines, find an anomaly, compare it with yesterday’s deployment, post a report, go back to sleep. Nobody has to call it.
Five layers make an agent operable.
Triggers wake a process. The process runs the agent. Governance bounds it, and management steers it from outside.
- 1TriggersAnything that can wake the process.
- 2ProcessA lifecycle, a context and a trigger queue around the agent.
- 3AgentThe Promptise Agent, with its memory and MCP tools.
- 4GovernanceBudgets, health, missions, the journal, scoped secrets.
- 5ManagementA REST API, a CLI, a dashboard and escalation.
Work starts the moment something happens.
Any trigger can wake the process, and events queue in arrival order. One process carries as many as it needs.
On a schedule.
Every five minutes, every Monday at nine, the first of the month. Standard cron syntax.
TriggerConfig(type="cron", cron_expression="0 9 * * MON")When a system calls.
GitHub pushes, payments, alerts. HMAC signatures optional.
When data lands.
Glob patterns on a folder, filtered by event type.
When another agent acts.
Topics with wildcards, inside the same runtime.
When work is handed over.
Decoupled coordination through a topic.
When you say so.
Register your own trigger type and use it like a built-in one.
Nothing is lost when a process crashes.
Every process keeps state with a full mutation history. A file journal and checkpoints bring it back after a restart.
# every process has an AgentContextctx = process.context # read and write state (timestamped, attributed)errors = ctx.get("error_count", 0)ctx.put("error_count", errors + 1) # long-term memoryhits = ctx.search_memory("previous database migrations")ctx.add_memory("v2 migration finished for orders") # full mutation history of a keyhistory = ctx.state_history("error_count")process stopped · restart_policy: on_failure
rebuilding from the last checkpoint
replaying the events after it
state, memory and counters restored
The agent picks up where it stopped. Secrets are never written to the journal, so none come back from it.
Autonomy your team can sign off on.
Budgets, health checks, missions and scoped secrets keep an agent inside an envelope you define.
An envelope the agent cannot exceed.
Limits on tool calls, turns, cost and irreversible actions, per run and per day. A payment can weigh more than a search. When it hits the limit: pause, stop or escalate.
health = HealthConfig( enabled=True, stuck_threshold=5, # identical calls in a row loop_window=6, error_window=10, on_anomaly="escalate", # log | pause | escalate)Problems caught without a single model call.
Stuck calls, loops, empty replies and rising error rates are spotted from behaviour. It is off by default and turned on per process.
An agent that works until the goal is met.
Give it an objective and success criteria. A judge checks progress every few runs, with limits on runs and time.
Credentials scoped to one process.
Each process holds only what it needs, for as long as it needs it. Secrets expire, rotate without a restart and live in memory only.
Steer agents while they run.
Adapt.
Fourteen meta-tools let an agent change its instructions, add triggers and spawn processes, within limits you set.
Talk.
Send a directive or ask a question. The agent picks it up between runs, with no restart.
Manage.
A REST API of 37 endpoints, a client, the CLI and one-file manifests run the fleet.
Define the whole agent in one file.
Validate the manifest before deploying, version it with your code, and talk to the agent once it runs.
# pipeline-monitor.agentname: pipeline-monitormodel: openai:gpt-5-miniinstructions: | Monitor data pipelines. Alert on anomalies. Cross-reference with recent deployments. triggers: - type: cron cron_expression: "*/5 * * * *" - type: webhook webhook_path: /alert webhook_port: 9090 budget: max_tool_calls_per_run: 10 max_cost_per_day: 50.0 on_exceeded: escalate health: enabled: true stuck_threshold: 5 config: restart_policy: on_failure max_consecutive_failures: 3from promptise.runtime import OrchestrationClient client = OrchestrationClient( "http://localhost:9100", token="...",) # tell a running agent something, without restarting itawait client.send_message( "incident-responder", "Ignore staging alerts, we're deploying", message_type="directive", priority="high", ttl=7200,) # ask it a question and wait for the answeranswer = await client.ask( "data-analyst", "What anomalies did you find today?", timeout=300,)Know which AI agent did what, and prove it.
“The AI did it” does not satisfy a security team. Every agent gets an identity on each call it makes and each action it records.
- 01
Local identity
A stable id, a name and an owner, asserted by the process. Free and instant, and enough to answer which agent did this, across traces, audit entries and journals.
AgentIdentity("billing-bot", owner="payments") - 02
Verifiable identity
Your identity provider mints a short-lived signed token. Resources verify the caller instead of trusting a name, with one credential per audience.
identity.get_credential("api://billing")
AgentIdentity.from_entra("billing-bot", resource="api://mcp")Source: IMDS, or the federated token file on AKS. The credential is short-lived and refreshed before it expires. AgentIdentity.auto() detects the provider.
Traces
Every model turn and tool call carries the agent id.
MCP calls
A signed credential goes to the servers it calls, per resource.
Approvals and audit
Requests and audit entries name the agent.
Runtime
Journals, escalations and messages between agents carry it.
Start with one scheduled agent. Grow into a fleet.
The Harness runs inside Kubernetes or your scheduler. It is the part that understands an agent’s state, cost and behaviour.
- 01
One cron agent
A single AgentProcess with a cron trigger that checks something and reports back.
AgentProcess - 02
Add persistence
Switch to the file journal so it survives restarts, and keep state across runs.
file journal - 03
Several agents
One AgentRuntime manages many processes. The event bus connects them, budgets stop surprise costs.
AgentRuntime - 04
Govern and scale out
Health, missions and scoped secrets, then several nodes with RuntimeCoordinator and manifests.
RuntimeCoordinator
One framework, three parts.
Build with the framework. Run it on the platform. Launch it with our engineers.
The Harness is open source under Apache 2.0. Run your first agent today. The platform and our engineers are there when you need more.