cmdev Academy — Honest AI & Cloud Security Writing

117 articles

The Three-Tier Agent Loop as a Three-Tier Attack Surface

The three-tier agent loop is a good architecture and a wide attack surface. Orchestrator hijack turns every worker brief into an attacker-controlled task. Swarm consensus manipulation poisons the stateless worker fleet from a single injection point. Judgment-chain compromise turns the advisor's critique into a rubber stamp. Part 2 covers the three attack surfaces the tier logic creates, the principal-of-record discipline that keeps the audit chain intact across the tiers, and the regulatory grounding under the EU AI Act Article 14 and NIS2.

July 20, 2026Samuel A.11m

Engineering

The Three-Tier Agent Loop: What's Right, What's Thin, and What's Missing

The three-tier agent loop — orchestrator plans, workers execute in parallel, advisor critiques at commitment boundaries — is directionally right and worth building. It is also thinner than the LinkedIn thread that made it famous. Part 1 of two covers what the design gets right, where self-blind verification hides the biggest failure mode, the missing tiers below the LLM that most posts skip, and when the pattern is genuinely worth the operational tax. Part 2 covers the three-tier attack surface.

July 20, 2026cmdev10m

Engineering

LLM Resilience Patterns, Part 2: Implementation — LiteLLM Router, Redis Circuit Breaker, and the Prometheus Alerts That Fire

Part 1 covered the six resilience patterns and when to reach for each. Part 2 walks through the implementation — LiteLLM Router (what it gives you and where it stops), the Redis-backed circuit breaker underneath, Tenacity with jitter, semaphore-per-feature bulkheads, the Prometheus alert thresholds that fire before the tickets arrive, and the six anti-patterns that turn the resilience stack into resilience theatre.

July 18, 2026Mayowa A.10m

Engineering

Your AI System Shouldn't Fail Because One API Failed — LLM Resilience Patterns, Part 1: Design

Every LLM provider fails eventually. Rate limits, timeouts, regional outages, quota exhaustion, model deprecations — the question is not whether they happen, it is whether your users notice. Part 1 of a two-part series covers the six resilience patterns every production LLM application needs, when to reach for each, what each costs, and the minimum viable stack. Part 2 covers the implementation and where LiteLLM Router stops carrying the load.

July 15, 2026Mayowa A.10m

Cloud Security

Data Lake Security: Bucket Policies, KMS Envelopes, and the Six Assumptions That Break in Production

The data lake is the least-audited surface in most organisations that have one. Bucket policies drift, KMS grants outlive the team that created them, cross-account trust widens quietly, and the six specific misconfigurations that keep showing up in incident reviews are the ones that decide whether the next data breach is preventable or a headline. This piece is the honest read on data lake security — the discipline underneath the AI pipeline everyone else writes about.

July 14, 2026Samuel A.11m

Engineering

Streaming Data at PII Scale: Kafka, Schema Registry, and the Encryption Trade-Off That Decides the Cost Model

Kafka at PII scale runs into three walls at once — governance (which topics carry personal data), encryption (broker-level vs field-level, and the cost math that decides between them), and audit (how do you prove three months later which consumer read which PII field). This piece is the honest engineering read on how a Bangalore data-platform team actually navigated all three, with the cost numbers and the Schema Registry patterns that made the difference.

July 14, 2026cmdev11m

Engineering

Governing the Agent — The Working Vocabulary, Part 3: What It Takes to Run in Production

Part 1 covered what an agent is. Part 2 covered how agents reach the world and each other. Part 3 covers what it takes to run one in production without surprising your compliance officer, your customers, or your on-call engineer. Seven terms: guardrails, policy layer, sandboxing, human-in-the-loop, agent identity, observability, evaluations. Each named as a discipline, not a checkbox.

July 12, 2026Mayowa A.10m

Engineering

The Healthcare Vocabulary for Agentic AI — Five Terms the Core Twenty Do Not Cover

The three-part core vocabulary covers what every agent team needs. Healthcare adds five terms of its own — the clinical decision support boundary the agent cannot cross, the pharmacist-in-the-loop discipline for anything that touches dispensing, the granular data residency at consent that governs cross-border clinical data, the FHIR-portable state that keeps the patient record cross-institution from day one, and the clinical-corpus eval that measures the agent against curated Nigerian primary-care scenarios. This piece is the second vertical extension, grounded in the Sirehs project.

July 9, 2026cmdev10m

AI Security

The Banking Vocabulary for Agentic AI — Five Terms the Core Twenty Do Not Cover

The three-part core vocabulary covers what every agent team needs. Banking adds five terms of its own — the deterministic ledger boundary the agent cannot cross, the reconciliation loop that closes every agent-initiated financial action, the real-time fraud overlay that scores agent actions on the same substrate as human ones, the regulatory reporting trail that surfaces to DORA, CBN CSAT, SAR/STR and equivalents, and the authorisation envelope the agent operates inside. This piece is the honest vertical extension the core series promised.

July 6, 2026cmdev9m

AI Security

Coordinating the Agent — The Working Vocabulary, Part 2: How Agents Reach the World and Each Other

Part 1 covered what an agent is. Part 2 covers how an agent reaches anything outside itself — tools, other agents, other systems. Six terms: MCP, A2A, Agent Protocols, Orchestrator, Multi-Agent System, Handoffs. Each carries a specific trust boundary that the previous generation of software did not have to answer, and the vocabulary is where the boundary either gets named or gets skipped.

July 3, 2026Samuel A.9m

Engineering

Building the Agent — The Working Vocabulary, Part 1: Foundations

Most teams are building agents faster than they are building a shared vocabulary for them. Ask five engineers on the same project to define "agent" and you get five answers. Part 1 of a three-part working vocabulary walks through seven foundations — the loop, tool use, memory, RAG, grounding, context engineering, and the pipeline that ties them together. Each term gets a use-it-when rule, the trade-offs, and the honest cost of building it right.

July 1, 2026Mayowa A.9m

Engineering

HTTP 402 Wakes Up: What It Actually Costs to Charge AI Bots Through AWS WAF

AWS WAF can now return HTTP 402 Payment Required to AI bots — a price manifest in the response body, paid in USDC via Coinbase's x402 protocol. The status code that has been reserved and unused since 1990 just woke up at production scale. This piece walks through the stack that actually carries it, the cost surface that does not appear in the AWS launch blog, the architectural shift it signals, and what the feature does not solve.

June 30, 2026Mayowa A.10m