01Week 1
Frame
Map the jobs worth automating, the decisions that must stay human, and the one metric that defines success.
AI product studio— Lisbon · worldwide · since 2015
Practice notes · 2026Intelligencethat acts,designed to be trusted.
Lopezi designs, builds and governs agentic products — from the interface people rely on to the protocols agents speak.
|
Scroll
— The shift
The model is no longer the bottleneck. Trust, legibility and cost are. The moment AI stops answering and starts acting, design becomes infrastructure — and that is exactly where we work.
57%
of organisations already run agents on multi-stage workflows.
Anthropic · State of AI Agents, 2026
1 in 5
companies has a mature governance model for autonomous agents.
Deloitte · State of AI in the Enterprise, 2026
70%
of teams running production agents in I&O will face a material incident from weak runtime controls by 2029.
Gartner · forecast
— Practice
Five disciplines for the agentic era.
01 / 5Agent Experience
Interfaces wherehumans and agentsshare the work.
When software acts on its own, the interface becomes a contract. We design how agents show their plan, ask for permission, explain a decision and hand control back — so people stay in charge without micromanaging.
AG-UIGenerative UIVoiceDesign systemsFigma → code
02 / 5Agentic Systems
Agents thatfinish the job,not the demo.
Multi-agent systems on production-grade harnesses: orchestration, tools, memory and state that survive real traffic. We ship coding, research, operations and customer-facing agents that complete multi-step work end to end.
ClaudeGPTGeminiOpen-weightLangGraphTemporalOpenTelemetry
03 / 5Protocols & Agent-Ready Web
Make your productspeak agent.
Agents are becoming your next customers and your next integrators. We expose products through MCP servers, A2A agent cards and WebMCP, make content machine-negotiable, and prepare checkout for agentic commerce.
MCPA2AAG-UIWebMCPOAuth 2.1x402
Run the agent-ready checklist 04 / 5Evals, Guardrails & Governance
Trust is anengineeringdiscipline.
Autonomy is only as good as the controls beneath it. We build what lets you say yes to agents: evaluation suites, policy enforced below the model, agent identity, audit trails and circuit breakers.
EvalsRed-teamingOpenTelemetryPolicy-as-codeAI ActISO 42001
05 / 5AI Economics & Performance
Intelligenceat the rightprice.
Token spend is only the visible part of AI cost. We route every task to the smallest model that passes the evals, cache and batch aggressively, and measure cost per outcome — not per call — so AI scales with margin.
RoutingDistillationPrompt cachingBatchEdgeFinOps
— Method
From first sketch to agents in production.
1/5
02Weeks 2–3
Prototype
Clickable agent flows with real models behind them. We test trust and recovery, not just usability.
03Weeks 4–10
Build
Production harness, tools, memory and interface — shipped in weekly increments you can use.
04Continuous
Evaluate
Eval suites become release gates. Every change is measured on quality, safety, cost and latency.
05Ongoing
Operate
Tracing, drift monitoring and model upgrades, so the system keeps improving long after launch.
— Proof
Trusted wherefailure is expensive.
A decade of enterprise product work for Lufthansa, Allianz and Siemens — now applied to systems that act on their own.
50+
Products shipped
11
Years building
15+
AI systems in production
99%
Uptime SLA
Case study · Lufthansa
A booking platform for millions of passengers.
From design system to production code for one of Europe's premier airlines: faster booking, fewer drop-offs and a foundation built to evolve.
40%
Faster booking
2.5M
Daily users
98%
Satisfaction
— Journal
Field notes from the agentic web.
Agent Protocols8 min read
MCP, A2A, AG-UI: The Three Protocols of the Agentic Stack
Agents need three open contracts to leave demos behind: tools (MCP), peers (A2A), and people (AG-UI). Here is how the stack fits together — and what to ship first.
AI Governance9 min read
Don't Ship Autonomy Without Evals
Agentic AI fails where governance stops. Written policies cannot catch a runaway tool call — only eval suites, release gates, and runtime controls can. Here is a practical stack.
AI Economics7 min read
Stop Counting Tokens. Start Measuring Cost Per Outcome
Token spend is the visible tip of AI cost. Routing, caching, and the right small model turn a €0.84 ticket into €0.11 — without lowering quality if evals hold the line.
— Next
Let's design whatyour agents do next.
A 30-minute call. We'll map one workflow worth handing to an agent — and what it would take to trust it.
Book a call or write to info@lopezi.com