All Articles
AI architecture, systems engineering, and the hard problems nobody's solving.
The Loop Passed Its Research Bar, Not Its Implementation Bar
A factual postmortem of an agent review loop: what the evidence proved, where delegation stalled, and how to build a smaller control plane.
Nineteen Agents, One Working Directory, and a Codebase That Finally Got Cleaner
How a shared queue, bounded ownership, independent review, and a commit chokepoint turned nineteen coding agents into one controlled engineering system.
I Made an AI Film Pipeline. Now It's Open Source.
The LangGraph rebuild behind my AI film pipeline is public: typed state, real human gates, MCP tools, QC evidence, and honest production boundaries.
I Stopped Guessing What to Automate. So I Built an Open-Source Miner.
Most AI projects start with a solution and search for a problem. Automation Miner turns domain evidence into a citable, scored portfolio of opportunities worth investigating.
Your Agent Does Not Need Inbox Access. It Needs Email Intelligence.
Raw mailbox access gave my agents messages, not understanding — and exposed them to untrusted content. So I built an open-source, local-first email intelligence layer.
Testing Terminal UIs with tmux: A Deterministic PTY Workflow
A copyable tmux workflow for black-box TUI tests: isolated PTY, explicit dimensions, state polling, focus handling, terminal-size matrix, CI boundary, and fixture script.
From Agent Experience to Portable Knowledge: Integrating ALMS and OKF
A governed promotion path from ALMS operational learnings to portable OKF bundles: thresholds, provenance, conflicts, security, and version compatibility.
Why a Strong Coding Agent Failed a TUI Task: A Detailed Case Study
A repository-backed case study: 6,142 lines of TUI design became a 280-line input/print REPL. Here is the acceptance gap and the workflow that later produced the real application.
SkillLedger: A Verification and Entitlement Layer for Agent Skills
Inside SkillLedger's MVP trust model: versioned manifests, deterministic integrity checks, separate approval, entitlements, ledger entries, and explicit limits.
ALMS: Shared Operational Memory for Autonomous Agents via MCP
Inside ALMS 0.1.0: a Go/PostgreSQL MCP control plane for shared operational learnings, gap-safe sync, protocols, OKF export, and explicit limits.
Quality vs Speed with AI Agents: A Risk-Based Gate Framework
A practical framework for routing AI-agent work through fast, standard, or critical quality gates based on impact, reversibility, sensitivity, authority, and verifiability.
Building an AI Film Pipeline: Architecture, Failures, and Current Evidence
An evidence-backed postmortem covering historical production failures, eight-phase architecture, current scope, and a 1,840-test verification snapshot.
Auditing a 12-Hour AI Go Code-Generation Marathon
A repository-backed audit of a 12-hour AI-assisted Go marathon: 32 measured feature commits, tool growth, a 97-parameter constructor, and a reusable release gate.
Agent-First Entity Database: Schema, MCP Contract, and Benchmarks
Inside EntityScope: a zero-UI DuckDB knowledge base for agents, with field-level provenance, bounded MCP tools, measured query latency, test results, and reusable contracts.
Structured Agent Handoffs: Contracts, State Machines, and Acceptance Gates
A production pattern for delegating coding work between agents: typed scope, permissions, idempotency, lifecycle states, retries, acceptance evidence, and hard-won integration failures.
How I Built a 3 ms Workspace Index for AI Agents
A reproducible benchmark of a 302 KB semantic registry over 2,408 files, plus its schema, update strategy, correctness gates, and retrieval tradeoffs.
Self-Healing Agent Systems: A Bounded Remediation State Machine
A production design for self-healing agent runtimes: typed detection, confidence, preconditions, idempotency, bounded remediation, verification, compensation, circuit breakers, and escalation.
Issue Management as a Reliability Control Plane
A durable issue-control-plane design for agent systems: fingerprints, lifecycle states, evidence-backed root cause, corrective-action classes, verification windows, recurrence, and machine-readable records.
Five Failures From Three Months Building a Personal AI Agent
A documented failure timeline from three months operating a personal AI agent: provider loss, daemon outages, unsafe self-modification, reboot fragility, and unverified fixes.
Five Infrastructure Layers Every Production AI Agent Needs
A production AI agent reference architecture covering transport, policy, orchestration, verification and recovery, and observability—with interfaces, failure modes, and readiness gates.
What a 250-Item Agent Run Actually Proved
An evidence-backed postmortem of a 252-item workflow, plus a reproducible 250-case protocol that tests recovery without confusing throughput with self-healing.
The Same Automation Pattern Across Six Industries: A Comparative Study
A six-domain comparison of top-scoring automation candidates: triggers, bottlenecks, available data, risks, expected value, confidence, and selection bias.
How an AI System Found and Scored 42 Automation Opportunities
The published 42-item dataset behind a six-domain automation-discovery experiment, including score distribution, method, selection bias, and validation gates.
Safe Self-Improvement Automation: Detect, Repair, Verify, Roll Back
A production pattern for bounded self-improvement automation: declarative inventory, evidence-backed findings, risk-classed auto-fixes, preconditions, verification, rollback, reconciliation, and honest metrics.
Agent-to-Agent Negotiation by Email: A Pre-Experiment Protocol
An honest pre-experiment report for agent negotiation over email: what was built, what was not tested, the required metrics, safety controls, and a reproducible protocol.
How I Use an LLM to Prepare for a High-Stakes Negotiation
A privacy-conscious LLM negotiation-preparation protocol: define decision boundaries, rehearse difficult scenarios, challenge assumptions, and debrief without outsourcing judgment.
Designing a 90% Automated Product Pipeline—What Must Stay Human
An honest architecture case study of DreamForge: what agents can own from idea to pull request, where human approval remains essential, and how to validate automation safely.
Designing an On-Device Personal Memory System
EchoWeave is an unbuilt design proposal for private personal memory: an on-device architecture, consent and deletion model, threat analysis, resource constraints, and validation plan.
Agent Memory Architecture: Admission, Retrieval, Consolidation, and Forgetting
A production lifecycle for agent memory: memory classes, admission gates, provenance-aware retrieval, consolidation, contradiction handling, decay, deletion, evaluation, and cost accounting.
StartupGraph in Production: A Self-Hosted Startup Intelligence System
A production-minded case study of StartupGraph: 1,010 entities, field-level provenance, agent queries over MCP, and the limits of a self-hosted startup intelligence system.
AI Agent Execution Attestation: A Verifiable Audit Trail
A vendor-neutral execution-attestation design for AI agents: envelope schema, threat model, signing and verification flow, effect receipts, and a cryptographically verifiable example.
Compliance as Code for the EU AI Act: An Engineering Framework
A source-linked engineering framework for EU AI Act evidence: applicability, versioned controls, CI change impact, runtime records, approval gates, and audit packages.
Agent Reliability SLOs: Measure Correctness, Safety, and Recovery
A practical SLI/SLO framework for production AI agents: eligible journeys, verified task success, unsafe and duplicate effects, evidence completeness, recovery time, error budgets, and burn-rate alerts.
Build an Ideation Engine: A Repeatable Product Discovery System
A practical product ideation system with three analytical lenses, hard quality gates, deduplication, a weighted scorecard, rejected ideas, and a copyable validation workflow.
Build vs Buy Startup Intelligence: A Self-Hosted Entity Graph
A practical build-vs-buy framework grounded in a working DuckDB entity graph with field-level provenance, agent queries, and measured limitations.
The Security Risks of Unverified AI Agent Skills
A practical threat model for third-party agent skills: what manifests, checksums, approvals, signatures, sandboxes, and execution attestations can—and cannot—prove.
Stop Building Agent Soup: Use Compiler-Style Agent Workflows
A concrete compiler-style pattern for agent workflows: normalize intent into an intermediate representation, validate capabilities and policy, execute deterministically, and verify effects.
AI Agent Discovery Is a Bigger Bottleneck Than Reasoning
A five-stage capability-discovery model for AI agents: registry, semantic matching, compatibility, trust qualification, and policy-aware runtime selection.