Speed where it’s safe, control where it matters.
AI agents in production, under control.
I put AI agents to work in production and keep them under control as they scale: agents inside your engineering workflows, and agent swarms run by people. 25 years in software engineering, most recently nearly six years leading a backend department.
- ~20 agent roles run under written policies, every commit audited against 35 rules
- 25 years in software engineering
- 10 live global titles supported
- 95% less time to provision an environment
- +25% continuous delivery KPIs, four quarters running
- Over $100K in cloud savings across two years
Two patterns
Agents inside your workflows
An event in an existing process, such as a merge request or a ticket, starts an agent or a team of agents, and the result comes back to the people who need it.
Proof: agentic code review in the delivery pipeline of around 80 microservices, used by engineers every day.
Agent swarms run by people
A coordinating agent routes the work, specialist agents do it, and people decide what is theirs to decide.
Proof: around 20 agent roles I designed and ran in a team of six people and around 250 agents at Web Summit Lisbon 2026.
What holds both together is control and measurement: what each agent may do, what actually stops it, and what it costs.
Agent swarms run by people
Web Summit Lisbon 2026
In November 2026, collaborating with the RiskMandate team, six people and around 250 AI agents built and ran a live event site, the PrompTugal.news newsroom and native apps: about 20,000 commits across 15 repositories in ten days. My lane was the technical side of the team’s site and its graphs and live views, with around 20 agent roles of my own (895 commits on the team’s site).
- A written behaviour policy for every agent, and each agent confined to its own files, enforced in CI
- A policy agent auditing every commit against 35 rules, each classed as enforced by a tool, checked after the fact, or rule only
- One secret per environment, each session signing its work with its own key, and my own go required for anything in my name, reaching outsiders, keys or irreversible
- Estimates against actuals, and cost, tracked for every piece of work
- The team’s command centre: a live map of hand-overs between people and agents, and a replay of every commit
In practice
Most of the cases below come from nearly six years leading the backend department at Outplay Entertainment, a Scottish mobile games company: around 80 microservices across half a dozen environments, behind its live global titles.
Agents inside workflows
Agentic code review, in production
A delivery partner built an AWS-funded prototype for agentic code review. I rebuilt it for our use case and put it into the delivery pipeline, where engineers used it every day. It coordinated multiple specialised agents in parallel, each covering a different review angle: style, quality, UI, security, and Continuous Delivery KPI measurement.
Four things made it work: configuration that could change as we learned, feedback loops engineers actually used, frequent iteration, and KPIs solid enough to tell us if it was working. I also worked with AWS and the delivery partner to structure the commercials so the business paid nothing: a $0 invoice.
Next: an agent team for code review that the delivery workflow starts through an API, designed for dynamic choice of specialist agents, replay-based evaluations and token budgets (in development, to be open-sourced).
Other prototypes
I also built an experimental tool: automated, multi-language, pre-emptive content moderation for player chat capable of detecting toxic and inappropriate language.
The rest of the platform
I stabilised legacy platforms until alert volume dropped from daily to yearly, letting us retire 24/7 on-call entirely, and drove the modernisation of half a dozen critical backend systems — LiveOps & their management tools, notification campaigns, leaderboards, the legacy game backend — without pausing support for live titles or new launches. I also built the backend function into a proper department: career paths, specialisation tracks, growth frameworks.
Working together
Hands-on engagements
Senior contract work building agent integrations, agent swarms and platform capability, at a day rate — direct with clients, or as a technical partner for agencies that want someone who has done this in production.
Fixed-fee diagnostic
About two weeks, ending in a written assessment of AI adoption readiness, agent controls or delivery performance. The right start if you have bought the tooling and cannot tell whether it is working, or are deciding whether to.
Fractional engineering leadership
A monthly retainer for teams that need senior engineering leadership without a full-time hire.
Also open to broader senior work: backend and platform engineering, AWS serverless and traditional compute, microservices at scale, legacy modernisation and DevOps.
Writing
- Set up your AI agents and decide what each may do: an interactive guide for executives, security, data protection and engineering readers
- Speed where it’s safe, control where it matters: how code checks and review share the work across 100+ agents
- Angelo’s communication graph: the newsroom’s piece on the command centre I built
- Ten days at Web Summit: Dinis Cruz’s account of what the team built
Built with open-source tools, credited in each article.
Get in touch
A short note about what you are dealing with is enough to start.
- Email angelo@leggiopasqua.com
- LinkedIn linkedin.com/in/angelo-leggio
- Call Book a 30-minute call