Product · Early access

Best Agents

The governance runtime for production AI agents.

A runtime for AI agents that companies can actually trust in production. Every action is identified, traced, and audited, and a person approves anything risky before it runs.

Request a demo See how it works

run 9c0bfafaacme · policy desk

14:02:11run.start agent=policy-desk user=u-1042 tenant=acmeIdentified
14:02:12tool.call company_knowledge_searchAllowed
14:02:14tool.call bind_policy POL-94874Waiting on you
14:03:40approval.granted by=underwriterApproved
14:03:41audit.append #201 prev=4e9e1201bca9…Signed

Chain verified · 201 events · no tampering

Sample data from the demo workspace.

Who it is for

Teams whose agents take actions, not just answer questions.

Insurance operations

Agents answer underwriting questions from approved documents and draft the paperwork.

Needs approval Bind, cancel, file a claim

Customer support

A support agent that resolves requests over the API, with every step traced.

Needs approval Refunds, account changes

Research

Agents that search, read, and summarise, citing where each claim came from.

Needs approval Paid tools, sending results out

How it works

Every call your agent makes goes through the runtime first.

Your app

REST API or streaming

Best Agents runtime

Identity: agent, user, tenant
Policy check on every tool call
Human approval for risky actions
Budgets, kill switch, eval gates

Out

Model provider

Tools over MCP

Your documents

Written for every step

Signed, hash-chained audit log · OpenTelemetry traces

Approvals

A person signs off before anything risky runs.

Binding a policy, cancelling a contract, filing a claim: you decide which actions need approval. The agent stops, shows what it wants to do and the policy that requires a check, and waits for someone to approve or deny it.

The exact action and arguments the agent wants to run.
The policy that requires a check, and when the request expires.
A second person must approve. Nobody can approve their own request.

An agent pauses before binding an insurance policy and waits for a second person to approve, showing the policy that requires the check.

Grounded answers

Answers from your documents, with every step checked.

Agents answer from the knowledge you have approved and cite where each answer came from. Every tool call is checked against policy, and the whole run is visible: messages, tools, plan, and errors.

The whole run: messages, tools, plan, and errors.
Each tool call checked against policy before it runs.
Every answer cites the document it came from.

An agent answers an underwriting question with cited sources, each tool call marked as allowed by policy.

Audit

An audit log you can prove nobody changed.

Every action is tied to an agent identity and written to a hash-chained log with signed checkpoints. Verify the chain in one click and hand auditors a report they can check offline.

Every hash and signature recomputed from the first event.
Merkle roots signed with the workspace Ed25519 key.
A report auditors verify offline with the public key.

The audit integrity screen showing a verified chain of 158 events and a signed checkpoint.

Screens from the demo workspace with sample data.

For developers

Three headers, one identity chain.

Call an agent over the REST API. The agent, user, and tenant you send are attached to every action that follows, so each step in the audit log says who asked for it.

OpenAI
Anthropic
OpenTelemetry (OTLP)
MCP tool servers
Postgres
Docker
Python

curl -X POST https://your-host/api/v1/agents/support/run \
  -H "Content-Type: application/json" \
  -H "X-Agent-Id: my-agent" \
  -H "X-User-Id: user-1" \
  -H "X-Tenant-Id: tenant-1" \
  -d '{"input": "Help me reset my password"}'

What ships today

Everything a production agent needs around it.

Agent identity

Every action carries who ran it, for which user and tenant, so nothing happens anonymously.

agent=policy-desk user=u-1042 tenant=acme

Kill switch

Stop an agent, or every agent, immediately. Revoke an identity without redeploying.

kill_switch.engaged runs stopped: all

Eval-gated releases

Agent and prompt versions are scored, and only promoted when they pass.

v13 score 0.91 >= gate 0.85 promoted

Tracing and logs

OpenTelemetry spans and structured logs on every run, sent to the tools you already use.

span gen_ai.chat 412ms 1,204 tokens

Cost controls

Token budgets and cost limits per run, with durable execution that survives restarts.

budget $0.50 spent $0.03 ok

Tools over MCP

Connect external tool servers through an MCP gateway, under the same policy checks.

mcp crm.lookup_customer allowed

Skills and domains

Versioned skills and YAML-configured agent behaviour per use case, no code plugins.

domains/support.yaml skill v4

Tenant isolation

Data, members, and knowledge are scoped per tenant, and every change is audited.

scope tenant=acme row-level security

Sample values shown under each capability.

Compared

Tracing shows what happened. This decides what is allowed to happen.

Capability	Best Agents	Tracing-only tools	Build it yourself
See every step of a run	Yes	Yes	You build it
Pause a risky action before it runs	Yes	No	You build it
Human approval with expiry	Yes	No	You build it
Signed, tamper-evident audit log	Yes	No	You build it
Kill switch and identity revocation	Yes	No	You build it
Evals that gate releases	Yes	Partly	You build it

Deployment

Runs where you need it.

SaaS: We run it for you.
Bring your own key: Our runtime, your model provider account.
Self-hosted: Docker on your own infrastructure.
Library: Embed the harness in your own Python service.

Not finished yet

What we are still building.

Automatic prompt and tool improvement is in progress.
The REST API is the complete channel today. Slack, Discord, email, webhook, and widget channels are early.
Knowledge search is lexical today. Hybrid keyword and vector search is next.

Questions

Before you ask.

Which models does it work with?

OpenAI by default, and Anthropic is supported. The provider and model are configuration settings. With bring your own key, usage runs on your own account.

Can our data stay on our own infrastructure?

Yes. Run it self-hosted with Docker, or embed it as a library inside your own Python service.

How does approval work?

You decide which tools need approval. When the agent calls one, the run pauses, shows the arguments and the policy that requires a check, and waits. Approvals expire, and undecided ones are denied.

Is it SOC2 certified?

No. It is built to produce SOC2-oriented evidence from day one: agent identity, OpenTelemetry traces, and signed audit logs. The certification itself is yours to run.

How do we connect our app?

Through the REST API today: run an agent, or stream its output over server-sent events. Slack, Discord, email, webhook, and widget channels are early.

What does it cost?

We are in early access. Request a demo and we will talk through pricing for your use case.

See it on your own use case.

Tell us what you want your agents to do, and we will walk you through it in a short call.

Request a demo Email us