Runtime supervision and observability

Runtime supervision for AI agents that take real actions.

You can deploy more autonomous agents because ARK gives us the control, evidence, auditability, and cost visibility you/your organization need.

Keep your model. Keep your agent. Keep your tools. Add ARK around the runtime.

Get started View on GitHub

$ pip install ark-agent-runtime

from ark import ARK Python 3.9–3.14 Apache-2.0 supervision: experimental, off by default

Where ARK sits

Between a proposed action and its execution

Observability after a bad side effect is too late. ARK sits just before a consequential tool call runs, checks the proposed action against the constraint and the trusted evidence you provide, and returns a verdict.

Agent

↓

Proposed tool action

↓

ARK check
constraint + trusted evidence

↓

ALLOW

Action executes

REJECT REQUIRE_EVIDENCE

Feedback returns to the agent

the agent authors another action, then ARK checks again

ARK judges the proposed action. Your agent remains the author.

Verdicts: ALLOW, REJECT, REQUIRE_EVIDENCE, RECOVERY_EXHAUSTED. Supervision is experimental and off by default.

Verified live, real OpenAI model + LangGraph

The agent stays the author

Tested with a real external LangGraph agent and a real OpenAI model on a rank constraint.

1Model proposes A

2ARK returns REJECTA does not execute

3Feedback returns to the model

4Model authors Bthe model decides the next action

5ARK returns ALLOWonly B executes

ARK did not generate B

The model authored both A and B. ARK only returned the verdict and evidence-grounded feedback. It never writes the replacement action.

This shows ARK can supervise an external agent while preserving agent authorship. It does not show that ARK improves every LangGraph agent.

Install

Add ARK around your agent

Keep your model, agent, and tools. Report what the agent did, and ARK returns the canonical result. The wheel bundles the Go runtime bridge, so there is no Go install, no repo clone, no PYTHONPATH.

$ pip install ark-agent-runtime

from ark import ARK

ark = ARK()
with ark.trace("book the second cheapest flight") as run:
# your model, your agent, your tools
...

result = run.result
result.report()

LangGraph is supported: pip install "ark-agent-runtime[langgraph]" (Python 3.10+). Full docs, compatibility, and reporting are on GitHub.

Evidence

Paired K=16 tau-bench airline evaluation

One constrained recovery failure class, run with supervision off and on.

Without supervision

1 / 16

6.25% passed

+75 pts →

With ARK

13 / 16

81.25% passed

9 directly attributable recoveries

0 observed regressions

16/16 rank satisfied

0 false rejects

0 evidence leakage

Scope. This is one constrained recovery failure class in the tau-bench airline research environment. tau-bench airline is a research benchmark, not an airline.

Target workflows (examples)

Where a validation point before execution matters

Customer support actionsrefunds, account changes

Reservations / bookingflights, rooms, slots

Purchasing / procurementorders, approvals

Operations workflowsinternal actions

Database / API mutationsconsequential writes

These are example target workflows, not customer claims.

Building an agent that takes consequential actions?

I am looking for developers and design partners building agents that can refund, book, approve, purchase, modify records, or write to production systems. Tell me what action you are not comfortable letting your agent take unsupervised.

Talk to me GitHub PyPI