Agent harness

The runtime between a model and your enterprise

Models are good at proposing the next step. The harness decides whether that step is allowed, runs it safely, checks the result, and records all of it.

01 · Model

Proposes a step

02 · Harness

Checks it against policy

03 · Harness

Runs the tool

04 · Result

Verified and returned

The model proposes. The harness decides.

1 loop

for chat, voice and background work

5 layers

from context to feedback

Any model

the rules don’t change

chat

voice

phone

background

Anatomy of a turn

One request, followed through the harness. Every stage is enforced outside the model, so a clever prompt can’t skip one.

“Summarize this account’s open questions and draft a follow-up.”

trace · turn_7f3a

stage 1/7

0ms

[id]

caller=user_214 role=field account=scoped

▇

Illustrative trace. Every turn gets one, replayable end to end.

Five layers, one core

Layer 01 · What the model sees

Working context

Account history, open questions and what worked before.

Account history

Context loaded for the caller’s account.

Open questions

The questions still waiting for an answer.

Prior sessions

Previous sessions inform the current turn.

Evidence

References travel with the answer.

Layer 01 · What the model sees
Working context
Account history, open questions and what worked before.
Account history
Context loaded for the caller’s account.
Open questions
The questions still waiting for an answer.
Prior sessions
Previous sessions inform the current turn.
Evidence
References travel with the answer.
Layer 02 · What the model can call
Actions
The model proposes a step. Tools run with scoped credentials; code runs in an isolated sandbox.
Plan
Propose a step before executing it.
Policy check
Rules are checked before anything runs.
Scoped tools
Credentials are minted per run.
Sandbox
Model-written code runs in isolation.
Layer 03 · What the harness enforces
Boundaries
Every tool call is checked against role, account and content rules before it runs. Enforcement lives in the harness, not in the prompt.
Identity set by the harness
The model never decides who the user is.
One policy layer
Role, account and content rules in one place, for every agent.
Safety paths
A possible adverse event switches the turn to the safety path and blocks other tools.
Scoped credentials
Minted per run, expired after it.
Layer 04 · What gets checked
Verification
Judges review the draft before it reaches the user.
Draft review
Check the proposed answer.
Citations
Return the answer with its references.
Before delivery
Nothing reaches the user unchecked.
Record
Save the result with the turn.
Layer 05 · How it improves
Feedback
The harness finds weak spots, drafts fixes, and tests them in eval worlds built from replayed workflows.
Watch production
Find weak spots in real workflows.
Propose
Draft a change for evaluation.
Test
Use replayed workflows in eval worlds.
Human approval
A person approves every release.
01
Talking with a person
Chat, voice and phone, in real time. Checks are sized to fit inside a spoken turn.
Latency budget per turn
02
Acting for a person
Copilots and assistants that work on someone’s behalf.
Sees the intersection of its scope and the user’s, never more
03
Acting alone
Watchers, monitors and scheduled jobs with no one in the loop.
Named owner · scope that expires · credentials per run

Generates and evaluates 
its own improvements.

The harness finds weak spots, drafts fixes, and tests them in eval worlds built from replayed workflows. A person approves every release.

Running · step 1 of 6

Illustrative improvement cycle

Step 1

Watch production

Watch the workflows running in production.

Signals and tools

Replayed workflows

Production signals

What the harness guarantees

Seven rules the harness holds, whichever model runs on it.
G1
Every tool call is checked before it runs
One policy layer evaluates role, account and content rules on each call.
G1
The model never sets identity
Identity is attached by the harness, never taken from the prompt.
G1
Model-written code can’t reach the internet
Code runs in an isolated sandbox, with no internet access.
G1
Nothing reaches a user unchecked
Judges review the draft before it reaches the user.
G1
Every turn can be replayed
The answer and citations are returned, and the whole turn is saved.
G1
A person approves every release
An engineer or reviewer reads the results, then approves or sends the change back.
G1
Models can change, rules don’t
The seven harness rules hold whichever model runs on it.
One loop · five layers · any model · a person approves every release
Run your agents on a harness
built for regulated work.
Talk to our engineers

Agent harness

The runtime between a model and your enterprise

Models are good at proposing the next step. The harness decides whether that step is allowed, runs it safely, checks the result, and records all of it.

See a turn run

01 · Model

Proposes a step

02 · Harness

Checks it against policy

03 · Harness

Runs the tool

04 · Result

Verified and returned

The model proposes. The harness decides.

1 loop

for chat, voice and background work

5 layers

from context to feedback

Any model

the rules don’t change

chat

voice

phone

background

How a turn runs

Anatomy of a turn

One request, followed through the harness. Every stage is enforced outside the model, so a clever prompt can’t skip one.

“Summarize this account’s open questions and draft a follow-up.”

trace · turn_7f3a

stage 2/7

0ms

[id]

caller=user_214 role=field account=scoped

38ms

[ctx]

loaded 3 sessions, 5 open questions

▇

Illustrative trace. Every turn gets one, replayable end to end.

The layers

Five layers, one core

Layer 03 · What the harness enforces

Boundaries

Every tool call is checked against role, account and content rules before it runs. Enforcement lives in the harness, not in the prompt.

Identity set by the harness

The model never decides who the user is.

One policy layer

Role, account and content rules in one place, for every agent.

Safety paths

A possible adverse event switches the turn to the safety path and blocks other tools.

Scoped credentials

Minted per run, expired after it.

01

Talking with a person

Chat, voice and phone, in real time. Checks are sized to fit inside a spoken turn.

Latency budget per turn

02

Acting for a person

Copilots and assistants that work on someone’s behalf.

Sees the intersection of its scope and the user’s, never more

03

Acting alone

Watchers, monitors and scheduled jobs with no one in the loop.

Named owner · scope that expires · credentials per run

How it improves

It proposes and tests

its own changes

The harness finds weak spots, drafts fixes, and tests them in eval worlds built from replayed workflows. A person approves every release.

Running · step 5 of 6

↰ Sent back · a person approves every release

Step 5

Person approves

An engineer or reviewer reads the results, then approves or sends the change back.

Signals and tools

Engineer review before release

Approval recorded with the version

Specification

What the harness

guarantees

Seven rules the harness holds, whichever model runs on it.

G1

Every tool call is checked before it runs

+

One policy layer evaluates role, account and content rules on each call.

G2

The model never sets identity

+

Identity is attached by the harness, never taken from the prompt.

G3

Model-written code can’t reach the internet

+

Code runs in an isolated sandbox, with no internet access.

G4

Nothing reaches a user unchecked

+

Judges review the draft before it reaches the user.

G5

Every turn can be replayed

+

The answer and citations are returned, and the whole turn is saved.

G6

A person approves every release

+

An engineer or reviewer reads the results, then approves or sends the change back.

G7

Models can change, rules don’t

+

The seven harness rules hold whichever model runs on it.

One loop · five layers · any model · a person approves every release

Run your agents on a harness

built for regulated work.