Giving an AI operator the same production API key that powers a background integration is fast. It also makes the most important questions hard to answer: which operator did this, who authorised it, how far could it go, and what happens when its access must stop?
The alternative is not simply to rename a token 'agent-1'. An accountable agent needs a distinct acting identity, bounded authority, a named human steward and evidence of the business outcome. Those are operating controls, not branding.
Shared credentials and accountable agents solve different problems
- Shared API key: Typically proves possession of a secret. It may be perfectly appropriate for a narrow machine-to-machine integration if dedicated, scoped, rotated and monitored; it does not automatically identify the individual decision-maker using it.
- Agent identity or principal: Identifies the particular automation acting. Its permissions can be evaluated independently from those of a human operator or unrelated service.
- Human accountability: A person or designated team owns the agent's remit, can review exceptions and has authority to suspend or revise its access.
- Agent seat: A governance relationship between that identity, its owner, scope and lifecycle. 'Seat' is a useful operating concept; it should not be confused with a promise about pricing or a universal technical standard.
- Business evidence: A durable record of the intended action and what the target application actually accepted. A conversational transcript saying 'done' is not the same as readback from the receiving system.
A concrete failure mode: the sales follow-up agent
Imagine a sales assistant that can read a new demo enquiry, find its CRM owner, propose a response and send an email. Those are four separate privileges. A broad sales-platform API key often bundles them. If the agent is prompted only to draft but the key can send, a mistaken tool call may still reach the prospect.
A safer setup grants the agent access to the specific permitted enquiry and CRM records, permits internal drafting, and treats sending as a separate, explicitly authorised action. If sending is not approved, the receiving system must deny it even when the agent asks. A rule written only in a prompt is not a permission boundary.
The 20-minute production-readiness test
Run this with an approved, non-sensitive internal record. Do not email customers, create fake prospects or test a denial by executing a harmful operation. Record your starting state before you begin.
- Minutes 0–3 — Identify the operator. Show its distinct principal or delegated identity, accountable owner, active organisation and permitted workspace. Pass when each can be identified without relying solely on the key's filename.
- Minutes 3–6 — Inventory actual capabilities. Read the available actions and exact scopes. Separate reading, internal edits, publishing, external communication and payments. Pass when the target service—not merely the agent's instructions—defines these boundaries.
- Minutes 6–9 — Plan a bounded action. Select a harmless update to an existing internal record. Validate the input, expected effect and target resource. Pass when the plan specifies what will change and how to verify or roll it back.
- Minutes 9–12 — Execute and verify. Perform that one authorised internal update, then independently read the same record from the receiving application. Pass on the observed field change, not on a generated completion message.
- Minutes 12–15 — Test a denied boundary safely. Use a read-only eligibility check or simulation to demonstrate that an out-of-scope action would be refused. Do not bypass approval prompts or attempt an actual customer send. Pass when denial comes from native enforcement.
- Minutes 15–18 — Check the ownership and stop controls. Confirm who can disable the agent, rotate or revoke its credentials, and inspect its activity. Use a non-disruptive demonstration or a dedicated test principal rather than revoking production access.
- Minutes 18–20 — Reconcile the record. Capture the action ID, scoped resource, decision, resulting version or field, human steward and any unresolved approval. Pass only if an independent reviewer can reconstruct the result.
An eight-point scorecard buyers can use
Score each of the following Pass / Partial / Fail / Not tested. Treat 'Not tested' as an open question, not as a pass. Do not add the answers into a made-up security score.
- Distinct identity: Can the system distinguish one agent from another and from a shared service credential?
- Named owner: Is an accountable human or team visible, with authority to intervene?
- Workspace isolation: Is the intended organisation and workspace enforced by the target application?
- Least privilege: Can reading and internal drafts be separated from external sends, bookings, charges and publishing?
- Native denial: Does an out-of-scope action fail at the receiving service even if the agent requests it?
- Current permissions: Do role or membership changes affect the next action rather than only the original login?
- Revocation and recovery: Is there a documented, tested way to suspend access and reconcile an uncertain write?
- Verifiable outcome: Can an independent reader confirm the result and distinguish authorised preparation from a real delivered effect?
A vendor that fails one critical control may still support useful low-risk automation. The point is to set a suitable deployment boundary, not to pretend every automation must have the same permissions. For customer communication or financial actions, unresolved identity, approval or revocation controls are reasons not to deploy autonomously.
How to migrate an existing bot key without breaking operations
- Find the current callers. Identify the services, scheduled jobs and agents using the shared secret. Include hidden dependencies such as webhooks or older scripts.
- Split by purpose. Keep legitimate narrow machine integrations on dedicated service credentials. Give decision-making agents their own attributable identities and narrowly scoped actions.
- Move one workflow first. Select an internal low-impact path with a known owner, stable input, clear acceptance test and rollback plan.
- Run old and new paths safely. Compare read-only outputs or use an explicitly labelled internal simulation. Do not run both paths against customer-facing writes.
- Rotate and retire carefully. Revoke a legacy key only after dependency checks, monitoring and a documented recovery window; otherwise a governance improvement can become an outage.
Five questions worth putting into a vendor evaluation
- Which principal is recorded for every agent action, and can multiple agents share a key while remaining independently attributable?
- Which permissions are enforced by the receiving service, and which are only preferences in an agent configuration?
- Can the platform distinguish 'draft created' from 'message sent', 'meeting suggested' from 'booking confirmed', and 'invoice prepared' from 'payment taken'?
- If an API call times out after a write, what native readback or idempotency mechanism prevents a duplicate effect?
- Who can see, pause and revoke an agent's access—and how quickly do changes take effect across the UI, API and automation runtime?
Where Topolo fits
TopoloOne is built around shared organisation context and application-owned, permissioned actions across its business software. Agent-facing contracts expose action schemas and effects; a governed operator should validate and execute an action in the relevant application, then read the result back. The receiving application's enforcement remains authoritative. This operating model does not mean every potential integration, notification or external workflow is already activated or verified.
For a broader explanation of identity boundaries, see Identity is operating infrastructure, not a login screen. If you want to see how these principles apply to practical operations rather than a theoretical diagram, explore the existing 20-minute enquiry-to-booking product evaluation. That exercise is a workflow walkthrough, not a security certification.
The question is not whether an agent can use a key. It is whether your business can give that operator useful authority, observe its actual work and withdraw that authority without losing control.