II Agent Harness

Secure. Deploy. Measure.

Bring Claude Code, Codex, OpenCode, or your own harness.
Keep your infrastructure. Add the building blocks around it.

Roadmap · Integration packages are planned.

A checkmarked access lock, retained deployment versions, and a connected run trace. — Secure accessVersion & deployMeasure the run

01 / Secure

Real access boundaries.
Before the agent runs.

Give agents the credentials and resources their job needs. Check the setup for indirect access paths and verify who can launch it. Your providers enforce the boundaries.

Example validation and authorization

Check first. Start only when authorized.
Protected operation	Gate result
Viewer DeploymentApply a new agent version	Denied
Operator DeploymentFresh required checks passed	Allowed
Required check UnansweredRequired checks must pass	Blocked
Agent A Agent BInvoke B’s configured setup	Denied

An unanswered required check does not count as a pass. Access checks happen before the operation starts.

You provide Provider access rules, verified identities, and credential delivery. A prompt is not an access boundary.

Permission integration

02 / Deploy

Version the whole setup.
Deploy it on your terms.

Keep the harness, container, repository, permissions, and environment together in a versioned setup. Deploy a known revision again without rewriting what earlier runs used.

Example versioned deployment

Configuration change

Research agent

v16 → v17

Changes from deployment 16 to deployment 17
Configuration	Before	After
Workload	digest a81c…	digest f09b…
Compute	2 CPU / 4 GiB	4 CPU / 8 GiB
Target	Local infrastructure	Local infrastructure

Setup 16Existing deployment

Setup 17New deployment

Next setupFuture deployment

A setup edit does not update the resources of an existing deployment.

Local and Google Cloud first. AWS, Azure, and GitHub are planned targets.

You provide Your harness, compute, provider access, and workload lifecycle. You start, track, recover, and tear down the work.

Deployment integration

03 / Measure

Understand the run.
Evaluate the work.

Link inputs, outputs, usage, and retained artifacts to the setup that produced them. Debug a run, compare changes, or evaluate the same outputs again without rewriting the original execution.

Example telemetry UI

Repository fix

Completed

Initiated by YouInvoked Agent B

Trace 4 spans

Agent B2.4s
1. Inspect repositoryModel · 600ms
  Starts at 0ms
2. Read sourceTool · 200ms
  Starts at 600ms
3. Prepare patchModel · 1.6s
  Starts at 800ms

Invocation details

Input: “Fix the failing test.”
Output: Candidate patch captured.
Evaluation pending.

Invocation: AuthorizedCheck record 08

Input tokens: 800
Output tokens: 200
Estimated cost: $0.004

Setup 17 · Deployment 16 · Custom harness 2.4.1

Example execution. Completion and evaluation are separate; one execution can have several grading attempts.

You provide Storage, retention, an evaluator, and task-specific grading inputs. SWE-bench and Harbor integrations are on the roadmap.

Telemetry integration

04 / Management

Reusable building blocks.
Your local control panel.

Use the local console to configure setups, review access findings, and inspect linked deployment, execution, and evaluation records. Extend the integrations for your own environment.

Example management UI

Workspace overview

Last 7 days

Invocations: 1,284
Allowed requests: 96%
Evaluation pass rate: 89%

Agent definition

Name / ID: Research agent / Agent B
Role: Configured invoker
Invocation authorization: Trusted identity required

Definition guide

Configuration

Deployment: Local infrastructure · v17
Capture: Redacted input + output
Export: Configured destination

Integration guide

Repository fix Run 042

View trace

Example records. Evaluation outcomes, deployment results, and lifecycle reports remain distinct.

You provide Your local environment and workload updates. Closing the panel does not stop the agent.

Local console integration

Make it part of your project

Bring your harness.
Keep your infrastructure.

Keep your models, identities, credentials, compute, and telemetry storage. We’re building the reusable infrastructure around them. Use the pieces you need; extend the rest.

Read the integration guide