Identity
Who is using the agent?
HARNESS ONE / ENTERPRISE
Move work forward.
Keep control of how it happens.
Define what agents can access, which tools they can use and when your team needs to approve. Build those boundaries into the loop where the work happens.
Discuss your requirementsRole
What responsibility does it serve?
Access
Which data and tools may it use?
Approval
What must be authorized before execution?
Action
What is permitted to execute?
Audit
What happened, and what was recorded?
01 / GUARDRAILS AND AUTHORITY
Know who can do what.
Define the agent’s authority in relation to the user, the business role, the tool and the action. Make the boundaries explicit before execution.
- Identity and role
Establish who is using the agent and which responsibility it serves.
- Data and tool permissions
Specify the information it may access, the tools it may invoke and the actions those tools may perform.
- Approval and stopping conditions
Identify decisions that require a person’s authorization, along with the conditions that pause or stop the work.
02 / CONTEXT, TOOLS AND MEMORY
Make the business explicit.
Connect the agent to the definitions, information and tools its role requires. Specify the task state it needs to retain as work progresses.
- Meaning and calculation
Agree business vocabulary, metric definitions and the calculation rules that approved tools must apply.
- Business data, prepared for the work
Identify the sources each agent needs. Align business definitions, validate data quality and freshness, and provide access through approved tools.
- Source ownership and exceptions
Define who maintains each source, how it is kept current, and how the agent should respond to missing or conflicting information.
- Connected tools
Map the sources, connectors and execution capabilities required for the agent’s scope.
- Memory and continuity
Define what task state must be retained, who may access it and how long it should remain available.
- Workflows and exceptions
Describe dependencies, handoffs and the route for questions or exceptions that the agent cannot resolve within its authority.
03 / EVALUATION AND OPERATIONAL EVIDENCE
Keep the outcome in view.
Test the agent against the work it is expected to perform. Use execution traces and observed outcomes to investigate failures and assess changes.
- Business scenarios
Define normal cases, edge cases and expected outcomes with the teams responsible for the work.
- Execution visibility
Follow tool calls, actions, approvals, results, quality, latency, cost and failures.
- Change and ownership
Review new versions, test for regressions and establish responsibility for ongoing operation.
THE OPERATING STANDARD
Reliability is built into the work.
Guardrails define permitted behavior. Verification checks individual results. Evaluation tests whether the agent performs as expected across business scenarios.
Governed
Permissions, policies and approval boundaries define how agents work.
Grounded
Business definitions, relevant data and operating rules inform each step.
Evaluated
Behavior is tested against business scenarios and expected outcomes.
Observable
Tool calls, actions, results, cost and failures can be traced and reviewed.
WORKING WITH YOUR TEAMS
Start with a defined operating scope.
Business and technical teams bring the requirements needed to configure the Harness around real work.
Map the work
Define the business role, tasks, expected outcomes, relevant information and systems.
Configure the Harness
Identify the loop, tools, context, memory, guardrails and evaluation scenarios the work requires.
Agree the operating requirements
Review deployment constraints, data handling, integration scope and responsibilities together.
BUILT AROUND YOUR BUSINESS
Bring your operating requirements.
Start with the work, the systems involved and the responsibilities you want agents to take on.