Custom Agents
AI Agent Operations
Operational support for agent workflows, with agreed monitoring, change and recovery responsibilities.
Available now · Defined controls · Action and outcome records
The operational problem
Handover is the beginning of an agent’s operating life
Models, APIs, credentials and business rules change after deployment. An agent can appear available while approvals are stuck, a connector is refusing unexpected calls or a backup has never been restored.
Agent Operations defines the support scope around runtime health, refusals, change and recovery. Owners, access, reporting, escalation and evidence requirements are agreed for the selected workflows rather than assumed from the build alone.
Operating scope
What has to be owned after launch
Runtime health
Availability, scheduled work, queue state and dependency failures are observed.
Approval flow
Waiting, rejected and expired approvals remain visible instead of looking like silent downtime.
Refusals and errors
Expected safeguards are separated from new failure patterns that require investigation.
Backup and recovery
The recovery design identifies which configuration, state and data need protection, who owns the recovery steps and what restore evidence is required.
Controlled change
Updates have defined tests, policy-based approval requirements and recovery options. Observation and update capabilities use separate permissions.
Operating method
How an agent remains observable and recoverable
- Define service ownershipEnvironment, contacts, support boundary, dependencies and escalation path are agreed during onboarding.
- Instrument the operating statesHealth, queues, approvals, refusals and external dependencies become observable signals.
- Back up what recovery needsThe recovery set and retention are defined from the agent’s actual state and configuration.
- Test restoreA backup is not reported healthy solely because a job completed. Restore evidence is recorded.
- Control and verify updatesChanges follow the agreed evaluation, regression, approval and verification requirements. Model, prompt, tool and knowledge-source changes are included in the change scope.
- Agree operating evidenceThe scope defines which activity, failures, refusals, changes and recovery evidence are reported, to whom and on what schedule.
Example report
Health, refusals and changes stay together
Illustrative monthly operations report - no customer runtime data
- Agents
- 3
- Uptime
- 99.9 %
- Refused operations
- 14 (all expected)
- Backups
- Daily, restore tested
- Updates
- 2 applied, approved
Reliability judgement
The rules behind managed agent operations
Healthy means the workflow can complete
A running process is insufficient if its queue, approval or connector dependencies are blocked.
Expected refusals are positive evidence
A blocked out-of-scope action demonstrates the boundary works; a new refusal pattern still deserves review.
Backups require restore evidence
Recovery confidence comes from a tested restore path, not a green backup icon.
Changes follow an agreed policy
Impact, reversibility and ownership determine the required checks and approval. Changes outside the accepted scope need a new decision.
Starting scope
What the operations onboarding delivers
- Ownership and escalation mapService owner, support roles, access requirements, escalation path and dependency responsibilities for the agreed scope.
- Monitoring baselineHealth, queue, approval, refusal and dependency signals configured for the agent.
- Backup and restore planRecovery scope, retention, procedure and evidence requirement documented.
- Update workflowTest, approval, deployment, verification and rollback stages agreed.
- Reporting specificationAudience, frequency, evidence sources and unresolved-state reporting agreed during onboarding.
Fit
When managed operations are or are not the right model
A good fit
- The agent was built or accepted with tests, a runbook and known operating boundaries.
- Your team needs defined responsibility for monitoring, change, recovery and escalation.
- The runtime can expose health, logs and controlled deployment operations.
Not the right fit
- The agent has no accepted specification, tests or ownership baseline.
- The expectation is unmanaged model changes with no approval or rollback.
- The environment cannot provide supported monitoring, backup or recovery access.
Control and approval points
Changes follow the agreed approval policy. Define the permitted operations and the action, approval and outcome records required for the workflow.
Operations questions
Questions a service owner should ask
What is agreed during operations onboarding?
The selected workflows, owners, access, monitoring, reporting, change, recovery and escalation responsibilities. The support scope is based on the existing system and operating requirements.
What counts as uptime for an agent?
The definition includes the workflow’s ability to accept work and reach its expected approval or result, not only whether one process is running.
How are refused operations reported?
They are grouped by expected policy refusal, invalid request and new anomaly so operators can distinguish a working safeguard from a developing problem.
Are updates automatic?
The change policy defines which operations may execute under pre-approved rules and which require a fresh decision. Relevant tests, evaluations, recovery options and verification requirements are agreed before execution, including emergency handling.
How often are restores tested?
The frequency is agreed from recovery requirements and operating risk. The report records the latest restore evidence, not merely the backup schedule.
Experience behind the service
Built from operating our own agent systems
The service uses the same operational disciplines TechOne applies to its own agent and product systems: deployment controls, monitoring, service management, backups, restore testing, TLS and approved updates.
Technical contact: David Máj, Founder & Technology Consultant.
Related architecture field guide