AI agents as teammates
Giving an agent a role name and goal does not make it a reliable colleague. When an agent can take multiple steps and use tools, role ambiguity creates action risk—not merely a poor answer.
The limit of the teammate metaphor
Thinking of an agent as a teammate helps design roles, service expectations, and handoffs. It does not give the agent personhood, equal status, or human accountability.
Agent Role Card
A goal says what the agent is trying to achieve; it does not define its authority.
Agent role contract
Goal ≠ Authority
- Goal
- Gather evidence on market options
- Inputs
- CRM · internal docs · market data
- Tools
- Search · Docs · Analytics
- May do
- Scan sources · compare · recommend
- Must not do
- Make investment decisions · share sensitive data
- Stop / handoff
- Conflicting sources · low confidence · high risk
- Human owner
- Strategy Lead
Seven parts of an agent role card
‘Do research’ is not a role definition. These boundaries need to be visible together.
Goal
Which outcome does the agent support, and how is success recognised?
Context
Which information does it need, and how does it judge freshness?
Tools
Which systems may it read or use?
Authority
May it only recommend, or also create, change, and send records?
Constraints
Which data, users, or actions are out of bounds?
Handoff
When does it transfer to a human, with which context and options?
Failure behaviour
Does it stop under uncertainty, use a safe fallback, and leave an event record?
Role-design case
A market-research agent with vague boundaries
An agent receives broad access to the internet, internal documents, and CRM. Its only goal is to 'find the best new market and accelerate progress.'
Four properties of sound delegation
Narrow but meaningful scope
The agent owns work connected to an outcome, with boundaries clear enough to prevent open-ended authority.
Least privilege
Tool and data permissions stay at the minimum scope required for the task.
Context-rich handoff
The agent reports what it tried, the evidence, and the human decision needed—not merely ‘I failed’.
Observable work
Actions, sources, errors, and intervention points can be inspected later.
Misconception
“An agent with a role and goal can be managed like a capable employee.”
Agents do not carry intent, social judgment, or organizational responsibility. Their performance depends on explicit boundaries, access control, evaluations, observation, and human ownership.
Take it to work
For one agent, fill in Goal / Inputs / Tools / Permissions / Stop / Handoff / Owner.
Be explicit about one prohibited action and the human role that takes over under uncertainty.
Private reflection · not saved
Keep this
Design the agent like a teammate, while keeping authority, observation, and accountability inside the human organization.