Custom Agents
AI Agent Build
A working agent on your process, delivered with tests and a runbook.
Available now · Defined controls · Action and outcome records
The operational problem
A working demo is not a system you can operate
An agent becomes operational only when its job, permitted tools, approval points, failure behaviour and acceptance criteria are explicit. Without those boundaries, a fluent answer can hide an untestable process.
We build from an approved process specification and tool contracts. The agent is tested against the scenarios you accepted, handed over with its code and runbook, and designed to refuse operations outside its allowed list.
Build scope
What a production-ready agent includes
Approved specification
The process, inputs, outputs, approval points and acceptance criteria are agreed before implementation.
Tool contracts
Each connected system exposes a short list of named operations with defined inputs, outputs and refusals.
Approval controls
Approval requirements follow the action’s impact, reversibility and your policy. Where approval is required, the exact action and target are visible.
Acceptance and evaluation
The agreed suite should combine tool and workflow tests with evaluation of decision quality, source evidence, uncertainty, refusals and failure paths.
Operational handover
Code, configuration, tests and a runbook explain how to deploy, observe, recover and change the agent.
Delivery method
How an approved specification becomes a working agent
- Freeze the accepted behaviourThe approved specification becomes the baseline for scope and acceptance, including what the agent must refuse.
- Connect only required operationsExisting or new connectors expose the minimum reads and writes the process needs.
- Implement the decision flowThe agent coordinates tools, evidence, state and human gates around the agreed process.
- Exercise failures as well as successesTests cover missing data, unavailable tools, denied operations and rejected approvals, not only the happy path.
- Assess acceptance evidenceResults are compared with agreed thresholds for task outcomes, decision quality, evidence, controls and failure handling. Open limitations remain visible.
- Hand over something operableYou receive the code, test suite and runbook, with a clear choice of who operates the agent next.
Example handover
Acceptance is visible, not implied
Illustrative acceptance handover - scope varies by process
- Scenarios
- 12 of 12 passed
- Approval points
- 2, working
- Runbook
- 9 pages
- Handover
- Scheduled
- Ownership
- Defined at handover
Engineering discipline
The rules behind a controllable agent
Permissions live in tools, not prose
A prompt saying “do not delete” is not a control. A connector that has no delete operation is.
Approval follows the action’s consequence
Where policy requires a reviewer, the approval covers the proposed action, target and relevant evidence rather than only a summary.
A refusal is a testable result
Out-of-scope and unsafe requests must fail predictably and leave evidence in the log.
Operations are part of the build
Monitoring, backup, recovery, configuration and ownership are documented before handover.
Changes need regression evidence
A change to the model, prompt, tools or knowledge sources should trigger the relevant evaluations and regression scenarios before release. Passing a previous suite does not establish that a changed agent is equivalent.
Starting scope
What the agent build delivers
- Working agentThe accepted process runs through the required tools and human approval points in the agreed environment.
- Tool permission mapAllowed operations, denied operations, credentials and approval requirements are documented per connector.
- Acceptance and regression suiteAgreed scenarios and evaluation criteria cover representative outcomes, boundary cases, evidence quality, refusals and failures, with rerun requirements for model, prompt, tool and knowledge changes.
- Audit and refusal evidenceThe handover shows what ran, what waited for approval and what the system refused.
- Code and runbookThe implementation, tests and operating instructions are yours rather than hidden behind a demonstration.
Fit
When a custom agent is or is not ready to build
A good fit
- One process already has an approved specification, accountable owner and testable outcome.
- The required systems can expose supported operations through APIs, connectors or controlled tools.
- Your team wants a maintainable implementation with code, tests and operating documentation.
Not the right fit
- The process changes with every person and nobody can approve a common operating rule.
- The request is still “build us an AI” without a named task, owner or acceptance criteria.
- Critical systems can only be accessed through unrestricted administrator credentials.
Control and approval points
You review acceptance evidence and open limitations. Define the permitted operations and the action, approval and outcome records required for the workflow.
Architecture and ownership
Questions a technical buyer should ask
What must exist before the build starts?
An approved process specification, named acceptance scenarios and access to the required tool connectors. If those do not exist, AI Process Design or a connector build is the correct first step.
Do we receive the source code?
Yes. The agreed implementation, configuration, tests and runbook are included in the handover. External platform licences or third-party services remain subject to their own terms.
Can the agent act without human approval?
Permitted actions and approval requirements are defined per workflow from impact, reversibility and organisational policy. Read access still requires authorisation and data controls. Where approval is required, it covers the exact action and target.
How do you prevent the agent from calling an unsafe operation?
The connector exposes only the named operations in its contract and rejects everything else. This boundary is tested independently of the model’s instructions.
What happens when a tool or model is unavailable?
The expected failure behaviour is part of the acceptance suite and runbook. The agent must stop or degrade visibly, preserve state and avoid repeating a write blindly.
What must be retested when an agent changes?
The change plan should identify which evaluations and regression scenarios must be rerun after a model, prompt, tool contract or knowledge-source change. It should compare the new results with the accepted baseline, including boundary cases and failures.
Who operates the agent after handover?
The operating owner, access requirements, monitoring, change and recovery responsibilities are agreed as part of handover. Any ongoing support is scoped separately.
Experience behind the service
Built the same way we build our own products
The build method comes from TechOne’s own product and agent work: multi-agent development roles, whitelisted system tools, cited research, and operational connectors for deployment, services, backups and TLS. We expose the architecture, tests and runbook without exposing customer systems or internal commercial data.
Technical contact: David Máj, Founder & Technology Consultant.
Related architecture field guide