Wiki

A map of ideas I keep returning to — mostly about AI, software and work. Each page is built from articles and talks, maintained together with an AI agent, and links back to its sources.

44 ideas · 35 people & organisations · 44 sources · A–Z

Recently updated: Agent-Mediated Markets, Agentic AI in Retail, Personal Agents, AI and Mental Effort

Engineering with agents

How software gets built when agents write much of the code: harnesses, loops, skills, reliability.

Harness EngineeringDiscipline of improving prompts, tools, context, hooks, sandboxes, memory, and feedback loops around AI agents.44 links in · 20 sources · in the blogAgent LoopsProgrammatic wrappers and loop-engineering systems that prompt, schedule, delegate to, evaluate, and supervise coding agents with durable memory…29 links in · 10 sources · in the blogAI-Assisted Software DevelopmentHuman-in-the-loop use of LLMs and coding agents for codebase research, implementation, tests, debugging, and tooling.33 links in · 18 sources · in the blog

• Modern Engineering Values — Agent-era software-engineering values: strong ownership, taste, strict guardrails, repo-local…
• Production LLM Reliability — Engineering patterns that make LLM systems bounded, observable, recoverable, evaluable, and…
• Loop-Dependent Software — Codebases and engineering workflows whose creation, review, diagnosis, or maintenance assume…
• Skillification — Workflow pattern for turning repeated agent tasks into durable, reusable skills that compound…
• Agent-Operable Environments — Task environments whose reset, rollout, verification, logging, safety, and permissions are…
• Agentic Software Factories — Software-production systems where humans direct architecture and acceptance criteria while…
• HTML Artifacts — Browser-viewable agent outputs for specs, reviews, reports, prototypes, and custom editing…
• Agentic RAG — Retrieval-augmented generation wrapped in an agent workflow with planning, tool routing…
• Agentic Code Maintainability — Code maintainability framed as an agent-navigation and token-footprint property, where clean…
• DRY Principle — Software-design principle that avoids duplicating knowledge, not merely similar-looking code
• Neurosymbolic Agent Systems — Contested framing of modern agents as symbolic orchestration around neural models, tools…

Thinking next to AI

What happens to judgment, effort and learning when the machine can do the thinking.

Cognitive SurrenderFailure mode where AI output replaces the engineer's independent judgment, creating cognitive and comprehension debt.29 links in · 12 sources · in the blogSpeedup IllusionMiscalibration where AI assistance feels easier and is expected to be faster, even when measured task time does not improve.9 links in · 1 source

• AI Job Grief — Psychological and social grief response when AI threatens professional identity, autonomy, and…
• Code Reading as Cognitive Labor — Reading and writing code as practices for building engineering mental models, not merely…
• AI and Mental Effort — How AI can polarize cognition by either replacing formative effort or amplifying human…
• AI Education With and Without Machines — UChicago curriculum sequencing: independent writing and reasoning before deliberate AI…

Work, teams and companies

Jobs, product work and organisational shape in the agent era.

AI Labor MarketEconomic and organizational impact of AI on work: lumpy displacement, job redesign, compute-constrained automation, transition insurance, and…15 links in · 8 sources · in the blogOrganizational MoatsDefensibility created by organizational shape, talent density, status, authority, and institutional design.15 links in · 7 sources

• AI-Native Product Work — Product-development pattern where cheap AI implementation shifts the bottleneck from building…
• Agentic AI in Retail — Retail shift from AI dashboards and recommendations toward governed agents that execute…
• Public Agent Collaboration — Practice of doing AI-agent work in visible, searchable team spaces so patterns, judgment, and…
• AI Usage Telemetry — Measurement of real-world AI diffusion through usage logs, artifacts, autonomy levels…
• Japanese Corporate Diversification — Japanese firms diversify because lifetime employment, horizontal coordination, internal…
• AI Outsourcing Transaction Costs — Coase/Williamson-style lens on whether AI lowers firm headcount through solopreneurship or…

Economy, compute and governance

Who owns the compute, who keeps the demand, and how the race could be slowed.

Dead Economy TheoryContested failure mode where AI erodes demand, ownership, purpose, and leverage, now paired with Restrepo's ownership-and-monetary-policy rebuttal.6 links in · 3 sources · in the blogAI Compute InfrastructurePhysical and market layer of AI access: GPUs, data centers, energy, inference budgets, compute opportunity cost, provider lock-in, and governance.12 links in · 4 sources · in the blog

• Test-Time Compute Evaluations — Model-evaluation frame that plots performance against inference tokens, cost, or time instead…
• Decentralized AI Compute — Open-protocol approach to AI compute where participants provide hardware and earn tokens…
• AI 2040 Plan A — Verified-slowdown scenario combining frontier-R&D transparency, distributed capability, staged…
• Agent-Mediated Markets — Consumer markets where personal agents research and buy for people, weakening firms' edge from…
• Mutually Assured Compute Destruction — Speculative treaty deterrent that makes strategic AI datacenters vulnerable to multiple powers…

Personal AI and knowledge

Agents that act for one person, software for an audience of one, and knowledge bases that maintain themselves.

Personal AgentsAI systems that move beyond chat to delegated online action while hiding technical setup from users.19 links in · 8 sources · in the blogLLM Wiki PatternPersistent markdown knowledge base pattern built from raw sources, maintained pages, and schema.6 links in · 1 source

• Self-Improving Knowledge Base — Knowledge base workflow where an agent maintains links, summaries, index, and lint health over…
• AI-Generated Personal Software — Malleable, AI-built applications for individual needs and tiny audiences, assembled from…
• LLM Wiki vs Obsidian — Tradeoffs between an agent-maintained LLM Wiki and an Obsidian-centered workflow.

Sleep, dreams and robots

Side quests: learning while asleep and robots that run their own experiments.

Sleep LearningModern evidence and ethical limits around learning, memory cueing, conditioning, and problem solving during sleep.2 links in · 2 sourcesPhysical AutoresearchAgent-driven real-world experimentation loop where coding agents improve robot policies through physical rollouts, automatic reset, and automatic…5 links in · 2 sources

• Lucid Dream Problem Solving — Use of lucid dreams for puzzle solving, skill rehearsal, and two-way communication with…
• Targeted Memory Reactivation — Sleep-research technique that replays cues linked to prior learning to bias memory…