ruicosta.ai
§ 00  ·  My place for ideas on artificial intelligence

My notes on AI adoption, Agentic Systems, and the teams of the future.

↓ scroll 01 · Opening
§ 01 · About
Portrait of Rui Costa

I’m Rui Costa. I’m passionate about amplifying ideas toward their maximum positive impact.

Fifteen years leading technology teams have shown me how much of our future is shaped by leaps of innovation that non-technical people struggle to see coming.

I work at the seam between deep technology and how organisations actually move. My path took me through distributed cloud systems, wireless networks, computer vision, and lately agentic AI.

This site is where I keep the ideas I share in keynotes and the systems I build. It isn’t a pitch. It’s a notebook kept in public.

§ 02 · Ideas

Some ideas on…

Three threads I keep returning to. On adoption. On building agents that do the work. On what changes inside a team when they do.

Part 01

AI adoption.

Four waves, two axes, and where most teams get stuck.

Adoption has a map. Two dimensions: inside versus outside the company, and today’s tools versus tomorrow’s work. That makes four quadrants. Most teams live in one. Three are wide open.

Today’s tools
Tomorrow’s work
Internal
Wave 01
Productivity gains
Internal × Today
Wave 03
Workflow redesign
Internal × Tomorrow
External
Wave 02
Customer experience
External × Today
Wave 04
New products & services
External × Tomorrow
Wave 01

Productivity gains

Internal × Today. The floor everyone stands on.

Faster emails. Meeting summaries. First-draft decks and reports. The easiest win, and the smallest prize.

It matters because it frees time. But if you stop here, you’ve only automated yesterday. This is where you learn the habit: drafts over decisions.

Wave 02

Customer experience

External × Today. Where the customer feels the difference first.

Faster responses. Better answers. Personalised service at every touchpoint. The quadrant with the most direct, visible payoff. People stay, refer, feel looked after.

It’s also where bad AI hurts you most publicly. The quality bar is high.

Wave 03

Workflow redesign

Internal × Tomorrow. The big one.

Agents owning end-to-end processes, not just assisting tasks. Fewer handovers. Fewer meetings. Fewer queues. The whole shape of how work flows changes.

Automate the repeatable so humans can amplify the valuable.

Wave 04

New products & services

External × Tomorrow. The imagination quadrant.

What can you offer a customer that simply couldn’t exist without AI in the loop? Not a faster version of today’s product. A new one.

The reframe question: “If we designed this from zero today, knowing what AI can do, what would it be?” Most teams miss this quadrant because it takes imagination.

Part 02

Agentic systems.

What works, and what doesn’t, once agents are doing real work.

These notes come from building and operating an always-on agentic framework inside a global company. Real workflows, real stakes, across engineering and analytics. Some of what I expected to work, did. Some of it really didn’t.

What worked

  • Orchestrator / worker separation. Deterministic code plans and dispatches. LLMs execute, under scoped tools. The orchestrator never implements.
  • Wave execution, not queues. Agents are API calls, not employees. Topological sort, group into waves, run the wave in parallel. Concurrency is effectively free.
  • Verify-reject-retry loops. A read-only verifier agent reviews every step. Rejections carry specific blockers so the retry is targeted, not vague.
  • Config-driven workflows. Declarative DAGs describe the steps. New workflows get added without touching the orchestrator.
  • Fresh context per step. No accumulated chat history. State moves between steps as structured data, not transcripts.
  • Role-scoped tools. Each role gets its own allowlist, enforced at runtime. Least privilege for AI, not just for people.

What didn’t

  • Vague delegation. “Make it look better.” Agents stall on under-specified asks. If it isn’t writable as a spec, it shouldn’t be delegated.
  • Delegating research. Discovery is better done directly with full tool access. Handing it to a scoped agent just rediscovers the obvious.
  • Over-packing a single wave. Too many agents at once means coordination overhead and context overflow. Break into phases; let the orchestrator do some work directly.
  • Equal-status agents with locks. Twenty peers contending for a lock do the throughput of two. Only hierarchical structures (planner, workers, judge) actually scale.
  • Wholesale framework swaps. Adopting someone else’s architecture wholesale to gain one feature is almost never worth the rebuild. Augment, don’t replatform.
Part 03

Agentic teams.

What changes inside the team, once agents are in the workflow.

The most visible shift is in handovers. From there it propagates through roles, meetings, and the shape of the org itself.

  • Handovers become artefacts, not conversations. When the next step might be an agent, the handover has to be written down. The spec is the interface. Standups get shorter, or disappear.
  • Review, not execution, is the scarce resource. Senior time moves upstream: less typing, more reviewing. The skill that gets rewarded is judging outputs quickly and well.
  • Observability replaces status meetings. If the system is legible (traces, logs, deltas, verifier rejections) the weekly update becomes a dashboard, not a meeting.
  • Scope expands per person; team size per initiative shrinks. A small team with good orchestration covers ground a larger team once needed. The org chart lags the leverage by a year or two.
  • Write it down, or lose it. If you can’t write the spec, an agent can’t run it. Honestly, a junior probably couldn’t either. The discipline pushes back into the human part of the process.
  • Taste, at the seams. Agents make more things faster. What stays human is the judgement at the edges: what to build, what to cut, what “good” looks like.
§ 03 · Keynotes

A log of talks, and what I said.

Recent stages. Each entry is the date, the headline, and the takeaways I actually wanted the room to leave with.

2026 · Feb

Fusing velocity and chaos to build the future of AI.

AI adoption is still microscopic, not saturated. Around 20% of businesses actually use AI versus 88% claiming they do. The window to win as an AI-native player isn’t closing, it’s barely opened. We are using the worst AI we will ever use.

Africa Innovators · Kigali, Rwanda

2025 · Nov

AI Fundamentals: adopting AI in your business.

A working stack for adoption, from data foundation to models to integration to governance to enablement. Where to start, what to skip, and how to sequence the layers so the company actually moves.

Blue Shuttle · EU project

2025 · Sep

Artificial Intelligence: the challenges of innovation and regulation.

We have to be AI-native within five years. Innovation and regulation are not opposites; the regulatory frame should make room for the experimentation that produces the next wave of products.

ANACOM Conference · Lisbon

2025 · H1

AI-driven innovation course.

AI compounds through data flywheels. The competitive shift is from economies of scale to economies of learning, and small focused teams with proprietary data and fast iteration now out-execute large incumbents.

Porto Business School

2024 · Dec

Behind the curtain: what wireless ad-hoc networks, AI, and e-commerce have in common.

Three careers in one stack. Vehicular mesh networks, AI mapping from dashcams, and a global roadtrip platform are the same problem in different costumes: orchestrating distributed nodes, real-time data, and compute at the edge.

University of Aveiro · keynote

2024 · Jul

Introduction to AI: pharma applications.

Where AI is mature enough to deploy in pharma today (literature summarisation, document workflows, customer support) versus where the regulatory and clinical bar still requires specialisation and human judgement in the loop.

IQVIA Conference

2023 · Dec

Intro to fundamentals of AI.

Short executive intro: what an LLM actually is, how to calibrate trust around hallucinations, and a first hands-on prompting block aimed at recruitment and operations workflows.

Mercan

2023 · Dec

Intro to fundamentals of AI.

Two-hour intro for a regional pharmacy network. Demystifying LLMs as prediction engines, where to trust them, and a first prompting practice block tailored to pharmacy operations.

Farmácias Silveira

2023 · Nov

AI introduction and panel.

For MBA students: the decisive AI skill for future leaders is fluency over delegation. Practising the tools yourself, not handing them to a specialist, is what will separate operators in the next decade.

Lisbon MBA · UNITE Summit

2023 · Sep

AI Bootcamp.

Same hands-on bootcamp format as Casais, run for the YPO chapter. Full day with executives moving through theory, live demos, and a first AI application built in the room.

YPO

2023 · Sep

AI Bootcamp.

Generative AI is a real step-change, not hype. The right executive response is to build personal fluency now and deploy inside the company, because the gap between organisations that adopt and those that wait is opening fast. Co-delivered with João Barros.

Grupo Casais

2023 · Aug

AI Bootcamp.

Same hands-on bootcamp format as Casais, tailored to the food sector. Traceability, documentation workflows, and customer-facing assistants came up as the most actionable starting points.

Grupo Mendes Gonçalves

§ 04 · Updates of the week

What I am reading, building, and noticing.

Short weekly notes. One signal, one takeaway, no filler. First entry coming soon.

Feed · pending · first entry coming soon

This slot will hold the most recent note. Archive and subscribe link will land alongside the first post.

§ 05 · Contact

Say hello.

Got an idea you want to kick around, a keynote to invite me to, or a team trying to figure out how to adopt AI well? Drop me a line.

hello@ruicosta.ai