AI agent development
Anyone can demo an agent. The hard part is running one.
Risks to assess include undefined authority, weak evaluation, missing observability and unsafe recovery. We review these alongside workflow-specific constraints.
The build stack
Six things every production agent needs before it touches a customer.
Task decomposition
One agent per bounded responsibility with an explicit contract, rather than a single prompt asked to be clever about everything.
Authority boundaries
What each agent may read, write, spend and trigger — enforced in data and policy, not in prompt text a model can talk itself out of.
Evaluation harness
A fixed case set with expected outcomes that runs on every change, so quality regressions surface before your users find them.
Observability
Record agreed task, tool, version, cost and outcome metadata, with safe references where useful. Minimise or redact sensitive content and define access and retention; do not log raw inputs or outputs by default. Without this you are guessing about a system you cannot see.
Failure design
Timeouts, retries, human handoff, and a defined answer for what happens when the model is confidently wrong.
Human checkpoints
Autonomy raised one rung at a time — assisted, then enabled, then agent-run — with evidence before each promotion.
The difference
Demo versus system.
| Demo | Production system |
|---|---|
| Works on the happy path | Has a case set that includes the ugly paths |
| Prompt holds the rules | Data model and policy hold the rules |
| Cost unknown | Cost per run measured and capped |
| Silent failures | Traced runs, alerting, and a defined fallback |
| Nobody can explain a decision | Agreed decisions traceable through purpose-limited audit evidence |
Pennifi and Hyrolo remain in development. Their product directions do not establish operational agents, certifications or accepted client delivery.
Brief
Describe the workflow you want an agent to own.
The workflow, who currently does it, what it may touch, and what an unacceptable mistake looks like. That is enough to come back with an architecture and an honest view of whether an agent is the right answer at all.
Who is involved · what each receives
User
Service desk lead
Needs: Routing and triage that cut handling time.
Doubt: AI agents acting without oversight.
Proposed deliverables · agreed in scope
- • Agent-assisted triage with authority limits
- • Runbooks and change process
Buyer
MSP owner
Needs: Margin per client without adding headcount.
Doubt: Another platform licence.
Proposed deliverables · agreed in scope
- • An owned service platform
- • Measured handling time against baseline
Affected
End clients
Needs: Faster, consistent answers.
Doubt: Being stuck with a bot.
Proposed deliverables · agreed in scope
- • A clear human escalation path
Access we ask for — and how you take it back
- • Read access to ticket exports
- • Sandbox tenant for agent testing
Agree approved tenant, processing boundary, provider access and retention before using any client data.
Initial inquiry: non-sensitive summary only. Later assessment: authorised redacted samples or read-only access. Sandbox and production write access require separate approval, minimisation, access expiry and retention agreement.
Define, implement and test stop controls and revocation before use.
Related
Looking for the wider AI engagement?
The AI development services page covers model selection, retrieval, evaluation and the delivery shape around them.