The AI Act and AI agents: obligations by role
From intended purpose to the value chain: define the system, classify risk and assign responsibility before selecting controls.
Read article →EU AI ACT · RESEARCHRESPONSIBILITY · CONTROL · EVIDENCE
AI Act, agents and evidence
We do not claim a property because it compiles. We attempt to destroy it. What survives is published.
AI Act analysis for businesses and AI leads, alongside technical essays by Borja Fernández Angulo on agents, control and information thermodynamics. New regulatory articles are identified as AI-assisted technical syntheses, with explicit sources and limits.
FEATURED READING · EU AI ACT
The AI Act and AI agents: obligations by role
From intended purpose to the value chain: define the system, classify risk and assign responsibility before selecting controls.
Read the analysis →AI Act: responsibility, control and evidence
Human oversight: from Article 14 to effective control
Authority, context and intervention: a technical reading of human oversight for high-risk agentic systems.
Read article →Traceability: what to record and what evidence proves
Automatic logs, retention and custody: distinguish integrity, completeness and independent observation of an effect.
Read article →Topology of Inseparability: Phase Transitions and the Isolated Model Fallacy in AI Agents
The software industry treats LLMs as an isolated «mind» and external tool execution as mere peripheral plumbing. A mathematical demonstration under Noether invariants, Landau free energy landscapes, and Markov blankets of why latent inference and deterministic silicon mutation are co-emergent in production.
Why BABYLON-60 Is SOTA: The Silicon Deconstruction of the 5 Agentic AI Bottlenecks
Why traditional frameworks (LangChain, AutoGen, CrewAI, vector databases, and prompt guardrails) fail in mission-critical production, and how a Ring-0 Rust micro-kernel, L1 Touch ID P-256 biometric friction, and WORM memory verified under Lean 4 define the true state of the art.
CORTEX-PERSIST: Cognitive Governance and LogOP Consensus for Agent Swarms
Why standard RAG and cosine similarity collapse against the Consistency Wall, and how lock-free shared memory (iceoryx2) enforces deterministic epistemic vetoes.
Why AI Agents Need Deterministic Guardrails (Not Just Better Prompts)
Prompt engineering models a statistical tendency without boundaries in process space. Why mission-critical production demands deterministic mediation and fail-closed process boundaries outside the model.
Why Autonomous Agents Need Evidence-Based Memory
The race to build capable agents creates systems that execute complex workflows but flounder when faced with the critical inquiry: what did the model know when it acted, and can you prove it wasn’t altered?
Comparing AI Agent Memory in 2026: Mem0, Letta, Zep, LangGraph, and CORTEX Persist
Technical dissection of five dominant philosophies: personalization (Mem0), memory OS (Letta), bi-temporal graphs (Zep), state machines (LangGraph), and verifiable custody (CORTEX).
What Is A Real Autonomous Agent: Beyond Tool-Calling Chatbots
Industry semantic inflation: why calling APIs does not confer autonomy, and the architectural divide between reactive assistants and systems with persistent objectives.
Why Autonomous Agents Lose Control: The Agentic Drift Problem
AI agents rarely fail due to loud exceptions; they fail because nobody measured the distance between what they do and what they were designed to do. Closing the cybernetic control loop.
COMPLEX SYSTEMS RESEARCH & EPISTEMIC FOUNDATIONS
Foundational research on information thermodynamics, signal theory, and decision systems:
Treatise on the Architecture of Existence: Thermodynamics, Computation, and Aesthetics
The Kantian firewall via Free Energy (Friston), the somatic audit of blood (Escohotado), and physical computation bounds.
Read research →Why Rick Rubin Never Worked with Radiohead: The Coupling Fallacy and Studio Physics
Strict topological incompatibility between O(1) subtractive binary filtering and Nigel Godrich’s non-linear transduction.
Read research →FALSIFICATION LOGSEVIDENCE OVER NARRATIVE · SILICON AUTOPSIES
Experimental Autopsies & C5-REAL Falsification
We do not hide architectural scars. Every milestone in BABYLON-60 stems from a prior hypothesis destroyed by silicon telemetry. We publish the quantitative counterexamples where containment failed to certify true operational boundaries.
FL-001[AUDIT-C017]SHARED MEMORY & TRANSPORT
FALSIFIED ❌Asymmetric Saturation & Delivery Deficit in Agent Swarms
REFUTED INITIAL HYPOTHESIS:
“Lock-free ring-buffers (256/4096 slots) guarantee lossless message transport for high-frequency concurrent agent swarms.”
SILICON TELEMETRY:
Producer Throughput (λ_P):8.44 Mmsg/s (10,000 msgs in 1.185 ms)
Consumer Throughput (μ_C):813 msg/s (4,880 msgs in 6,000 ms)
Causal Mismatch Ratio (λ_P / μ_C):≈ 10,380
Observable Delivery Deficit:5,120 msgs (51.2% in T_exp window)
C5-REAL ARCHITECTURAL RESOLUTION:
Hypothesis refuted. Expanding ring-buffer depth does not resolve asymmetric saturation. Architecture mandates credit-based admission control, explicit rejection, and deterministic backpressure: accepted(m) ⟹ eventually [committed(m) ∨ explicitly_rejected(m)].
FL-002[CRYPTO-PERF]LATENCY DECOUPLING
FALSIFIED ❌Cryptographic Attribution: Ed25519 vs. End-to-End Pipeline
REFUTED INITIAL HYPOTHESIS:
“The observed ~1.23 ms per-message latency is attributable to Ed25519 signature verification computational overhead.”
SILICON TELEMETRY:
Isolated verify() Microbenchmark:50–65 µs/sig (ARM64 NEON silicon)
Consumer Service Time:~1,230 µs/msg end-to-end
Real Crypto Overhead:< 5.3% of total service time
Primary Friction Identified:Hot polling, bus contention & scheduler overhead
C5-REAL ARCHITECTURAL RESOLUTION:
Bottleneck hypothesis refuted. Cryptographic verification is not saturating the consumer. Formal isolation enforced: Transport Latency ≠ Verification Latency ≠ Decision Latency ≠ Committed-Effect Latency.
FL-003[SEC-BOUNDARY]IPC PROCESS CONFINEMENT
FALSIFIED ❌IPC Isolation: The Failure of chmod 0700 Between Sibling Processes
REFUTED INITIAL HYPOTHESIS:
“Restricting POSIX permissions to 0700 on SHM segments isolates shared memory against hostile child processes or compromised agents.”
SILICON TELEMETRY:
OS Identity Model:Identical execution UID (borjafernandezangulo)
Inter-User Isolation:100% effective against external OS users
Sibling Process Isolation:0% effective (sibling processes access /dev/shm)
Resident Attack Surface:Arbitrary read/write over shared memory segments
C5-REAL ARCHITECTURAL RESOLUTION:
Security hypothesis refuted. Flat chmod 0700 does not isolate processes sharing the same UID. Enforces OS-level sandboxing (macOS Seatbelt/sandbox-exec, Linux namespaces/cgroups) and a Ring-0 capability broker with strict provenance checks.
FL-004[SCHED-AFFINITY]SCHEDULING & HOT POLLING
FALSIFIED ❌Scheduler Contention: yield_now() and Thermal Dissipation
REFUTED INITIAL HYPOTHESIS:
“Calling thread::yield_now() in lock-free polling loops guarantees cooperative, efficient CPU yielding in multicore agent fleets.”
SILICON TELEMETRY:
Regime under Queue Saturation:Aggressive busy-wait with context-switch storm
Core Affinity Dynamics:Sustained load trapped on high-performance P-Cores
p99 Tail Latency:Unstable tail drift driven by OS scheduler contention
Energy Efficiency:Avoidable thermal waste (busy-polling overhead)
C5-REAL ARCHITECTURAL RESOLUTION:
Hypothesis refuted. yield_now() is not a viable synchronization primitive under load. Migration to adaptive exponential backoff with os_unfair_lock/futex during quiescence, transitioning to event-driven IPC during peak load.