Digital workforce · machine actors

If agents become workers, they need defined authority.

The useful enterprise analogy is not a humanoid employee. It is a machine actor: software that can receive a goal, access data, use tools, communicate with other agents and sometimes change business state.

Human organisations already govern identity, role, ownership, permissions, delegation, records and revocation. As AI systems gain agency, equivalent control concepts need to exist for the software actors operating inside the company.

Explore the Agent Passport Understand agentic AI

Identity · ownership · permissions · delegation · expiry · revocation · audit

Machine actor record

AP-042 · Commercial Review Agent

Architecture concept

Owner

Commercial Director

Tenant

Example organisation

Purpose

Prepare variation evidence

Status

Active · bounded

Data

Project + contract records

Tools

Search · cost model · draft

Action ceiling

Prepare only

Approver

Commercial Manager

Authority state

Read ✓Prepare ✓Recommend ✓Issue notice —Approve cost —

First principle

Human users and machine actors are not the same identity problem.

A human signs in, understands organisational context and can be held directly accountable. A software workload may run continuously, hold credentials, make dynamic decisions and operate faster than any human supervisor can manually inspect.

Microsoft describes workload identities as identities assigned to software workloads so they can authenticate and access resources. SPIFFE similarly treats workload identity as a first-class security primitive. AI agents add a further governance question: not merely who is calling?, but what purpose, authority and decision boundary has this machine actor been granted?

The future company may have more machine actors than human users. They cannot be invisible.

Identity before autonomy

Identity

Which exact machine actor is this?

A stable identifier tied to a known trust and tenant boundary.

Authority

What may this actor access and do?

Data, tools, actions, delegation and thresholds expressed outside the prompt.

Accountability

Who owns the actor and its outcomes?

A named business owner, human approver, review cycle and reconstructable evidence.

An API key proves access to a service. It does not define business authority.

TEMRIK architecture concept

Interactive explorer

Inspect Maya's Agent Passport.

Maya is an illustrative Commercial Analyst. Select any passport field to see what it means, why it matters, what failure looks like and how the control should work.

Selected control

Agent ID

A stable identity for the machine actor.

What it means

Identifies one machine actor distinctly from users and services.

Why it matters

Without it, activity can be hard to attribute.

Example

Maya-COM-01

What failure looks like

activity can be hard to attribute.

TEMRIK control

Identity registry and workflow binding.

TEMRIK architecture concept · not a claim of current deployment

The Agent Passport.

A governance record that binds a machine actor to purpose, ownership, permissions, delegation boundaries, human authority and lifecycle state. It complements runtime identity; it does not replace an identity provider, OAuth, workload identity or A2A security.

Agent ID

Stable machine identity

Owner

Accountable human or function

Tenant

Organisational boundary

Purpose

Approved business objective

Model

Runtime model or provider class

Allowed data

Permitted information domains

Tools

Approved capabilities

Delegation rights

What work may be handed off

Action ceiling

Maximum consequence without approval

Human approver

Escalation authority

Expiry

When authority must be renewed

Status

Active · paused · revoked

Audit history

Material identity and authority events

Runtime identity

Prove which workload is calling.

Workload identity technologies can authenticate software without treating it as a human account. Patterns such as managed identities, federated workload identities and SPIFFE short-lived credentials reduce dependence on long-lived embedded secrets.

Governance identity

Prove why that workload is allowed to act.

The business still needs an explicit record of owner, approved purpose, data scope, tools, delegation, action ceiling, approver, expiry and status. This is the additional control problem the Agent Passport is designed to make legible.

Ownership

Every agent needs an accountable owner.

Ownership is not the same as authorship. The engineer who built an agent may not be the business person who should approve its purpose, data access or operating threshold.

Business owner

Owns the business outcome and acceptable operating boundary.

Technical owner

Owns runtime, integrations, reliability and identity implementation.

Data owner

Approves access to governed data domains.

Human approver

Holds decision authority for consequential exceptions or releases.

Security owner

Defines credential, access and incident requirements.

Reviewer

Periodically re-validates the passport and continued need.

Credentials

Do not confuse the agent with its credentials.

Credentials are evidence used to authenticate or obtain access. They should be scoped, rotated, revocable and—where infrastructure permits—short-lived or federated instead of permanently embedded in prompts, source code or agent memory.

Higher operational burden

Long-lived secret

Avoid where a stronger workload-identity pattern is available.

Platform-managed credentials

Managed identity

Useful when the hosting platform and target resource support it.

Trust without duplicated static secrets

Federated identity

Useful across CI, Kubernetes and cross-cloud scenarios.

Portable workload identity

SPIFFE / SVID

Useful where cryptographically verifiable workload identity and trust domains fit the architecture.

Permissions and action ceiling

Permission should describe more than a tool list.

A machine actor may have permission to call a finance system but still lack authority to release a payment. TEMRIK separates capability from consequence through the idea of an action ceiling.

Observe

Read approved information

No external state change

Prepare

Draft, classify, assemble evidence

Human releases the work

Recommend

Propose a decision or next action

Named human approval

Act

Execute low-consequence approved actions

Bounded by policy and thresholds

Escalate

Stop and request higher authority

Required for exceptions and high consequence

Tool access answers “can it call this?” Action authority answers “may the business allow this outcome?”

Delegation

Delegation cannot create authority from nowhere.

Multi-agent systems may hand tasks to specialists. That handoff should not silently grant the receiving agent broader data, tools or action rights than the originating business authority allows.

Agent A

Authenticated actor

Delegate

Defined task

Agent B

Known identity

Policy

Intersect rights

Result

Return + evidence

A safe default is delegated authority = the intersection of the delegator’s remaining authority, the delegate’s own passport and the policy for the requested task.

Agent to agent

Discovery is not trust. An Agent Card is not an approval.

A2A Agent Card

The A2A protocol uses Agent Cards to advertise an agent’s identity metadata, capabilities, endpoint, skills and authentication requirements. It relies on standard web security mechanisms for authentication and leaves authorisation to the implementation.

Enterprise policy

The company still decides whether this agent may be discovered, invoked, trusted for a specific skill, supplied with a particular data class, delegated a task or allowed to trigger an external action.

Identity enables interaction. Policy grants authority.

Expiry and revocation

Authority should decay unless deliberately renewed.

Time expiry

Passport requires renewal after a fixed review period.

Owner change

Pause when the accountable owner leaves or changes role.

Purpose change

Re-authorise when the business objective materially changes.

Permission change

Re-test when tools, data or action rights expand.

Model / runtime change

Review when a material technical assumption changes.

Security event

Immediate pause or revocation where compromise is suspected.

Revocation should address more than a UI status. Depending on the architecture it can require invalidating credentials, terminating sessions, disabling service identities, removing tool access, blocking A2A invocation and preventing queued work from releasing.

Supervision

Supervise the material events, not every token.

A useful digital-workforce control model records operational evidence: identity, goal, data boundary, tool use, delegations, policy decisions, approvals, exceptions, released actions and outcomes. It does not require exposing private model chain-of-thought.

Passport created

Credential issued

Permission changed

Task delegated

Policy exception

Human approval

Material action

Authority revoked

See the wider TEMRIK AI security architecture for identity, tenant boundaries, tool control and audit design.

Lifecycle

An agent should have a lifecycle, not just an API endpoint.

Register

Create a unique machine identity and bind it to a tenant, owner and approved purpose.

Attest

Establish that the running workload is the workload that should receive the identity.

Authorise

Grant the minimum data, tool, skill and action rights needed for the role.

Activate

Permit operation only after required policy, test and human checks are satisfied.

Supervise

Observe tool use, delegation, exceptions, approvals and material outcomes.

Review

Reconfirm purpose, owner, permissions, credentials, model assumptions and action ceiling.

Expire

End authority automatically unless a defined renewal condition is met.

Revoke

Disable identity, credentials, sessions and downstream access when risk or ownership changes.

TEMRIK control model

Machine actors belong inside the control plane.

The TEMRIK AI control plane is the architectural layer intended to keep company policy, identity, playbooks, tools, human authority and evidence around the models and agents doing the work.

Company policy

Purpose · risk · decision rights

Agent Passport

Identity · owner · tenant · lifecycle

Data + tools

Least privilege · scoped capabilities

Agent runtime

Model · orchestration · delegation

Dispatcher gate

Allow · restrict · approve · escalate

Human authority

Named approver · exception owner

Audit

Identity · action · outcome · review

Research & standards

Primary sources behind this architecture.

The sources below describe workload identity, agent-specific identity, A2A discovery and agent runtime control patterns. TEMRIK does not imply partnership, certification or implementation parity with these organisations.

NIST