Gehan Panapitiya Contract and advisory · agents in production · Christchurch, New Zealand

Contract and advisory · remote from New Zealand · any cloud or model stack

Agents are easy. Agents that stay stable in production for a year are not.

I keep AI agents stable, governed and secure in production, and I give an independent read on agent systems other people built. The method took an estate serving 3,000 corporate customers to agents closing 15% of incidents with no person in the loop, safely: low-severity incidents, through runbooks the team kept improving. Agents write the code; I write the harnesses, evaluations and guardrails that decide whether it ships.

If instead you are one person building a company that agents will run, and nothing exists yet, the last section is for you.

Tell me what is breaking See what I take on ↓
Validate · Review · Govern · Secure · Improve Embedded, scoped, or fractional

What I doFive kinds of engagement, each with a named deliverable

Validate

An independent read on an agent system you built or were sold.

Independent validation
Method: I state the threat model and the claims under test first, then run the system against a harness I write for it, and report what held, what failed, and what could not be determined. Deliverable: a written report you can show a board, a regulator or a customer, with the harness so the result can be re-run.
Independence
I do not validate a system I was paid to build or helped build within the previous twelve months, and I disclose any other interest before we start. A validation report is a review with its scope, method and limits stated, not a certification against a standard. Where a client needs a second reviewer, I name one and they sign it with me.

Review

What your agents and models are actually doing, and where the effort should go.

Usage review
How your team uses agents and models today: cost, drift, duplicated effort, quiet failures. Deliverable: the optimisations, ranked by payoff and effort, with the evidence for each.
Value discovery
Where AI would pay off in your business and where it would not. Deliverable: a short list of candidates with the trust and data each one needs before it can run.

Govern

Make trust in agents explicit, graded, and able to grow.

Incident and runbook review
Your last twenty incidents, sorted by which an agent could close safely through scoped tools and which must stay with a person, and why. Deliverable: the route to the first tranche of incidents closed without a page, and the trust grade that lets the tranche grow.
Governance and readiness
Which workflows AI can be trusted on today, which it cannot, and the controls that move the second group. Deliverable: a capability grade per workflow, the agent-identity and data-sovereignty controls to install, and the adoption plan that gets engineers onto the safe path.

Secure

Know your AI attack surface, and close it faster than it opens.

AI security posture
Where agents, tools and model access widen your exposure, and what is reachable today. Deliverable: the exposure map and the controls, in the order to install them.
Self-updating defence
A loop that watches published breaches and advisories, matches them against your estate, and aims to patch or block the same day, within a scope agreed up front. Built inside your stack with your team, so it is yours when I leave. The pattern is the one that contained a self-propagating worm in the AI package ecosystem on the estate I ran.

Improve

An improvement loop tuned to how your company actually runs and makes money.

Self-improving operations loop
I map one operational or value pipeline end to end, then design the loop that improves it from its own results: what it observes, what it may change, how improvement is measured against your numbers, and the separate checks that stop it when it gets worse. Deliverable: the loop design, its guardrails and measures, and a first version running on one pipeline, so the gain is shown on your data before it spreads.

How we workThree shapes. We pick one in the first conversation, and I say what it will cost before anything starts.

Scoped problem, returned

You hand me a defined problem; I take it away and return a result with the evidence and the next step. Typically two to six weeks. Fixed price.

Embedded, short term

I join your team for a fixed period, work inside your systems, and leave the capability with your engineers. Typically six to twelve weeks. Day rate.

Fractional lead

One or two days a week steering your platform or AI effort, without hiring a full-time head before you need one. Monthly, three-month minimum.

It is me doing the work, not a team I hand it to. I work remote from New Zealand and keep overlap hours with your time zone; for US and European teams that means early mornings and late evenings on my side, agreed up front.

For foundersBuilding a company that agents run

If you are one person with an idea for a company where agents do the operating and you hold the direction, nothing is breaking yet, and the offers above are not for you. This one is.

What goes wrong in a one-person company is the same thing that goes wrong in a 300-person one, earlier: nobody can say which work the agents are trusted on, the agents forget what they learned last week, and the one human becomes the bottleneck for every decision. Those are problem 1, problem 4 and problem 7 on the research page. The engagement brings twenty years of operating judgement to them now; the research covers them once problems 1 to 3 have results.

The engagement is a scoped one: three to four weeks, fixed price. Deliverable: an operating model for your company. Which functions agents hold and which you hold; how trust is graded and expanded as the record builds; the interface where your judgement enters cheaply and binds; and what must never be forgotten. You leave with a design you can build against. If you want me to stay on while you build it, that is the fractional shape above.

The same engagement applies to an existing service you want to re-deliver with agents running it end to end.

Tell me what you want agents to run

RecordThe estate the method was built on

15%
of production incidents closed by agents, no person in the loop (low-severity, runbook-driven)
60%
faster incident triage from cross-system correlation
30%
less alert noise from mining alert history into signatures
40%
engineering adoption of AI tooling in 16 months, by golden paths not mandates

Self-reported results from the estate I ran, delivered by the team I led; not independently audited.

Until October 2026 I led platform engineering for a business-intelligence SaaS, running the entire production estate for 3,000 corporate customers and 45,000 daily users: 30 cloud accounts, six Kubernetes clusters, four database engines, identity, observability, disaster recovery and the AI serving layer, with a six-person senior team and a $3M budget. The remediation agents act only through scoped tools and managed runbooks; destructive operations are out of their reach. When a self-propagating worm hit the AI package ecosystem, the standing controls, entry-point scanning and an allow-list enforced through a local proxy, contained it.

Before that: Kubernetes and AI operations at a global accounting platform, including a cross-model incident operator; five years of site reliability at 4,000 deployments a year for a board-governance platform; an eleven-engineer platform team for six embedded radio product lines; internal platform tooling at a global bank. Twenty years across business intelligence, financial software, board governance, embedded hardware and banking.

My independent research, on what a production agent should keep and forget, is at gehanpanapitiya.com. Engagements are where those problems get met in other people's systems; nothing observed inside a client's system becomes research data without written agreement.

Earlier recognition: gold medal, National Best Quality Software Awards (NBQSA), 2008; special recognition, Asia Pacific ICT Alliance Awards (APICTA), 2008; Microsoft Excellence Award, 2009; silver, Microsoft Imagine Cup, 2007.

StartOne conversation to pick the shape

Tell me what is breaking, or what you want agents to run. I reply within one working day, New Zealand time, and the first conversation is free.

Email request@gehanpanapitiya.com

Or message me on LinkedIn.