Insights

Poseedores de conocimiento.

Puntos de vista, investigación y pensamiento abierto de un equipo que lleva una década operando IA en producción y ahora construye agentes gobernados donde equivocarse es ilegal. Sustancia, precisión, honestidad intelectual.

StrategyAug 25, 20265 min read

Why the Giants Are Giving Agent Governance Away for Free

Microsoft, Google and Anthropic gave away the protocols and the governance machinery for AI agents, then placed them under one foundation. Free is not charity. It is how power scales.

GovernanceAug 25, 20265 min read

A Free Governance Toolkit Is Not Governance

The engine that enforces a policy is now free. The policy is not, because a policy is judgment. Why installing a governance toolkit is the start of the work, not the end.

GovernanceAug 25, 20265 min read

What a Governance Toolkit Cannot Do (From Its Own Docs)

A runtime governance layer for agents is necessary and not sufficient. Read the limitations, honestly stated, and the layers you still have to build around it.

ArchitectureAug 25, 20265 min read

Is MCP Just a Sexier Way to Call an API?

At the wire, an MCP server often wraps an API that already existed. The revolution is not the plumbing. It is that the reader changed, from a human programmer at build time to a model at run time.

ArchitectureAug 25, 20264 min read

MCP and A2A Are Not a War. They Are Two Floors of One Building.

MCP connects an agent to its tools. A2A connects an agent to other agents. The versus everyone argues about is a category error, and the real story is the stack forming underneath it.

StrategyAug 25, 20265 min read

When Your Agent Starts Hiring Other Agents

A2A lets an agent delegate work to a counterparty it did not build and may never have met. That turns using software into transacting with strangers, and identity becomes the hard problem.

ComplianceAug 15, 20266 min read

The EU AI Act Just Deferred Its Hardest Obligations. Do Not Exhale.

Brussels pushed the AI Act's high-risk deadlines to 2027 and 2028. The deferral buys calendar time, not readiness, and the pilots that died never failed on paperwork.

SecurityAug 15, 20266 min read

OWASP Says Prompt Injection Is Unsolved. Configuration Will Not Save You.

If the top LLM risk is structurally unsolved at the model layer, the only honest answer is a structural boundary, not a better classifier. Here is what that means.

SecurityAug 15, 20267 min read

The Research Is Converging: Agents That Think Must Never Act.

Independent 2026 research keeps reaching the same conclusion: the component that reasons over untrusted content must not be the component that acts. The reader needs no hands.

ComplianceAug 15, 20266 min read

DORA's Enforcement Phase Wants Evidence, Not Policy PDFs.

Supervisors have stopped accepting policy documents and started demanding evidence of resilience. An agent that cannot reconstruct what it did cannot pass a DORA review.

StrategyAug 15, 20266 min read

The Open-Weight Frontier Just Changed the Build-vs-Buy Math for Regulated AI.

Frontier-class open weights arrived in 2026. Sovereign, in-perimeter deployment is now a real option for regulated data, which turns model routing into a live decision.

SecurityAug 15, 20267 min read

The Agent Security Stack: Why No Single Control Holds

Defense-in-depth against indirect prompt injection. The four independent layers, why spotlighting alone is never enough, and how each OWASP ASI risk maps to a named control.

SecurityAug 15, 20266 min read

Non-Human Identity: Why Your Agents Need Their Own Credentials

Agents must not borrow human tokens. Short-lived least-privilege workload identity, egress allow-lists, and why a stolen agent credential should be worth nothing.

SecurityAug 15, 20266 min read

The Model Context Protocol and the New Tool-Security Problem

MCP gives agents hands, and hands change the threat model. Tool poisoning, the confused deputy, over-broad scopes, and why typed capabilities behind a PDP are the fix.

ComplianceAug 15, 20266 min read

Cryptographic Audit Trails: Evidence a Regulator Can Inspect

How to make an agent decision inspectable after the fact: append-only, hash-chained evidence on WORM storage, written before the decision commits, fail-closed.

GovernanceAug 15, 20267 min read

Evals and the Trust Threshold: Measuring an Agent You Can Defend

Evals are the mechanism that gates an agent's autonomy. What to measure, why one test is not an eval, and how thresholds move an agent up the autonomy ladder.

GovernanceAug 15, 20266 min read

Hallucination Is a Governance Problem, Not Just a Model Problem

In high-consequence work the answer to confabulation is not a better model but governance: grounding, abstention, concordance evals, and a human gate on the decision.

ArchitectureAug 15, 20266 min read

RAG Is Not Governance: Retrieval in Regulated Environments

Retrieval grounds an answer in your documents. It does not make that answer compliant. In a regulated environment, RAG needs the same governance as any agent action.

ArchitectureAug 15, 20266 min read

Fine-Tune, RAG, or Prompt: the Data-Governance Tradeoff

Three ways to give a model your domain, three very different data-governance profiles. Where regulated data lives decides auditability, erasure, and update cost.

DeploymentAug 15, 20266 min read

Open-Weight Models on Your Own GPUs

In a regulated context, an open-weight model inside your perimeter can beat a frontier API. Not because it is smarter, but because your data never leaves.

DeploymentAug 15, 20267 min read

Sovereign and Confidential Deployment: VPC, On-Prem, Air-Gapped, and TEEs

Deployment topology is a data-sovereignty decision, routed by sensitivity and failing closed. Where VPC and air-gap run out, confidential computing and TEEs begin.

EngineeringAug 14, 20266 min read

Shadow IT Doesn't Scale Into a Regulated Enterprise

A laptop automation and a production agent are separated by a chasm that no amount of prompting closes. The gap is architectural. Here is what it takes to cross it.

StrategyAug 14, 20267 min read

The Pilot Graveyard: Six Reasons Agentic Projects Get Cancelled

Around 40% of agentic-AI projects are projected to be cancelled by 2027. Most fail for the same six reasons, and each one is avoidable if you design for it up front.

GovernanceAug 14, 20266 min read

Your IT Governance Was Built for Software That Doesn't Change. Agents Change Daily.

Change-control designed for deterministic software does not hold on probabilistic agents. Graduated autonomy and evals are the governance agents actually need.

ArchitectureAug 14, 20266 min read

The Context Layer Nobody Built

Agents are only as good as the context they can reach, and most organizations never built the interoperability and identity layer agents need. Here is what that layer is.

ComplianceAug 14, 20266 min read

Compliance as Executable Policy

The EU AI Act and DORA are not documents to file. They are obligations that have to run inside the system as code, emit verdicts, and leave evidence a regulator can inspect.

DeploymentAug 14, 20266 min read

There Is No Off-the-Shelf Stack for Regulated AI

In a regulated enterprise, which models you can run and where they execute is a compliance decision, not a vendor default. Provider-agnostic, deploy-anywhere, sovereignty-routed.

Book a meeting

¿Quieres profundizar en alguno de estos temas? Tráenos el proceso que estás sopesando y te daremos una lectura franca de dónde encajan los agentes.

Reservar una reunión