TechOneDigital Talk to David

Custom AI agents

AI agent identity, guardrails and a kill switch.

An AI agent should not run under a person's account or a shared service account. Give it its own identity, put guardrails where it reads, decides and acts, and keep a tested way to stop it that revokes its access, not just its current task.

Written by · Founder & Technology Consultant

TechOne position

Treat the agent as its own actor.

An agent is neither a person nor a classic integration. It decides its next step at run time, from instructions and from content it reads, and it can be steered by that content. Access designed for a person or a fixed script assumes behaviour the agent does not have.

So the agent gets its own name in your identity system, its own narrow permissions and its own record of what it did. Then every action can be traced to the agent, to the person it acted for and to the approval it relied on, and the agent can be switched off without locking anyone else out.

Identity

Why the usual accounts fail for agents.

01

Running under a person's account

What it looks likeThe agent signs in as the employee who set it up and inherits everything that person may do.

What goes wrongA person's access is broad because people use judgement. The agent receives that breadth without the judgement, and the logs cannot tell what the person did from what the agent did. When the person leaves or changes role, the agent breaks or keeps access it should not have.

TechOne recommendation

Give the agent its own identity. When it acts for a person, record both: the agent as the actor and the person it acts for.

02

Running under a shared service account

What it looks likeOne technical account with wide rights is reused by several integrations, scripts and agents.

What goes wrongEvery agent can do what the most demanding integration needed. Nobody can say which agent changed a record, and stopping one agent means stopping all of them. A leaked credential exposes everything the account reaches.

TechOne recommendation

One identity per agent and per environment, with only the operations its process needs, and credentials that expire.

03

Acting for a person, or acting as itself

What it looks likeAn assistant drafts replies for one employee, while a background agent processes a shared queue overnight.

What goes wrongThese are different authorities. Acting for a person should never exceed what that person may do. Acting as itself needs its own approved scope. Mixing the two lets the agent borrow rights it was never given.

TechOne recommendation

Decide per process which mode applies. Token exchange standards separate delegation from impersonation; prefer delegation, so tokens show both the agent and the person.

Guardrails

Put guardrails where the agent reads, decides and acts.

TechOne recommendation: one filter on the model's input or output is not enough. Each control point below catches a different failure, and each has an owner.

01

What reaches the model

What it coversWhich data the agent may read and send to a model, and which models may receive it.

Why it mattersRead access is also access. Content the agent reads can carry instructions that try to redirect it, so retrieved text is treated as data, never as permission.

TechOne recommendation

Limit the records and fields the agent can read, check sensitive data before it leaves, and keep approved models on a list.

02

Which tools and records it may use

What it coversThe operations the agent can call and the records each operation may touch.

Why it mattersPermission to call a tool is not permission to change every record behind it. OWASP names excessive functionality, permissions and autonomy as the typical root causes of excessive agency.

TechOne recommendation

Expose named operations instead of a whole API, and check the object-level permission in the system that owns the record.

03

Which actions need approval

What it coversPayments, deletions, external messages and changes that are hard to reverse.

Why it mattersAn approval that says "allow the agent to update invoices" approves too much. The MCP specification says there should always be a human in the loop with the ability to deny tool invocations.

TechOne recommendation

Approve the exact action: this operation, these values, this record, for a limited time. A changed parameter needs a new approval.

04

What is recorded

What it coversThe evidence that lets you reconstruct why an action was allowed.

Why it mattersAn HTTP log shows that a call happened, not who asked for it, which rule allowed it or who approved it.

TechOne recommendation

Record the agent, the person it acted for, the operation, the policy decision, the approval and the outcome, including every refusal.

Kill switch

A kill switch revokes access, not just a task.

Stopping a running conversation is not enough. MCP cancellation asks a server to stop one request, and a server may ignore it, for example when the work has already finished. A real kill switch removes the agent's ability to act and then checks what already happened.

  1. 01

    Who may pull it

    Named people in operations and security, plus the business owner of the process. Pulling it must not need the team that built the agent.

    TechOne recommendation
  2. 02

    What it stops

    The agent's credentials are revoked, its tools are blocked at the gateway or server, and queued or scheduled actions are held. Other agents and people keep working.

    TechOne recommendation
  3. 03

    What happens to unfinished actions

    Actions that may have completed are reconciled against the system of record before anything is retried or reversed. A timeout is not proof that nothing happened.

    TechOne recommendation
  4. 04

    How you know it works

    It is tested before production and after major changes: pull it, confirm the agent can no longer act, confirm the record shows who pulled it and why, then restore access deliberately.

    TechOne recommendation

TechOne recommendation

Before the agent reaches production

  • The agent has its own identity in each environment, with an owner.
  • Its permissions list named operations, not a whole system.
  • It is decided whether it acts for a person or as itself.
  • Actions with consequences need approval of the exact action.
  • Every action and refusal is recorded with the person and the approval behind it.
  • The kill switch has named owners and passed a test.

Primary sources

Standards behind the protocol statements.

Statements about MCP, delegation and excessive agency link to their official sources. Everything marked TechOne position or recommendation is our architecture view.

  1. MCP authorization: security considerationsServers must only accept tokens issued for themselves; token passthrough is forbidden.
  2. MCP toolsThere should always be a human in the loop with the ability to deny tool invocations.
  3. MCP cancellationCancellation of an in-progress request, which a server may ignore in some cases.
  4. RFC 8693: OAuth 2.0 Token ExchangeDelegation versus impersonation when one party acts for another.
  5. OWASP Top 10 for LLM Applications 2025: Excessive AgencyExcessive functionality, permissions and autonomy as the root causes.

Direct answers

Questions teams ask before an agent goes live.

What is agent identity?

The agent's own account in your identity system, separate from people and from other integrations. It carries the agent's permissions, its credentials and its audit trail, so actions can be traced and the agent can be stopped on its own.

Can an agent simply use the employee's sign-in?

For a personal assistant that only drafts for one person, delegated access can work, if the tokens show both the agent and the person and the scope is narrower than the person's. An agent that acts on shared work should never borrow a person's account.

Are guardrails the same as a content filter?

No. A filter on prompts and answers is one control point. The guardrails that matter for consequences sit at the tools, the records, the approvals and the evidence, outside the model.

Does MCP give an agent an identity?

MCP defines how a client gets a token for a specific server reached over HTTP, and servers must reject tokens issued for someone else. Who the agent is, and what it may do in your systems, is still decided by your identity system and the systems of record.

Is a kill switch the same as cancelling a request?

No. Cancelling asks for one request to stop, and the work may already be done. A kill switch revokes the agent's access, blocks its tools and holds its queued actions, then reconciles what completed.

Where should we start?

With one agent and one process. List the systems it touches, whether it acts for a person or as itself, the actions that need approval and who may stop it. That list becomes the specification.

A useful place to start

Decide who your agent is before it acts.

Bring one agent or one process. We map its identity, its permissions, the actions that need approval and who can stop it, before it reaches production.

Review one agent with us