AI Agents

Software systems that wrap an LLM with tools, memory, and skills to autonomously complete multi-step tasks.

Last updated: 2026-04-13

Overview

An AI agent is more than a chat interface to a language model. It is an LLM embedded in a harness that gives it the ability to take actions, remember state, use external tools, and delegate to sub-agents. The harness is intentionally kept minimal; domain intelligence lives in reusable skill files rather than hard-coded scaffolding.

The field is moving fast: what requires elaborate scaffolding today gets absorbed into the base model with each new generation. Good agent design anticipates this by keeping instructions in readable markdown (easy to delete) rather than tangled code.

Core Concepts

ConceptSummary
coding-agentSix-component harness: context, cache, tools, compaction, memory, subagents
thin-harness-fat-skillsKeep the harness minimal; encode domain intelligence in skill files
claude-code-skillsNine skill types, writing tips, and production distribution patterns
model-context-protocolOpen standard giving agents access to external tools and data sources
agent-sandboxPer-user isolated execution environments for safe arbitrary code execution
context-normalizationConverting heterogeneous source data into uniform agent-readable format
s3-first-architectureS3 as durable source of truth with PostgreSQL as a query-optimized read layer
agent-evaluationDomain-specific eval suites and eval-driven development for production agents
agent-first-softwareSoftware designed for agent consumption: 100–1000x agent volume, semantic selection, enterprise diffusion gap
agent-memoryThree-layer memory model: MEMORY.md, episodic-memory (vectors), QMD (on-device search)
auto-researchAutonomous agent loop that improves a system without human involvement; program.MD; SETI@home pattern

Key People

  • andrej-karpathy — LLM Wiki pattern; auto research; macro actions; model jaggedness; loopy era framing
  • sebastian-raschka — six-component coding agent framework
  • garry-tan — thin harness, fat skills architecture pattern
  • nicolas-bustamante — production agent architecture in financial services (Fintool)
  • paul-hoekstra — Agentic Engineering four-layer framework (Configuration, Capability, Orchestration, Guardrails)
  • aaron-levie — Box CEO; agent-first software design; agent-as-extension-of-self
  • marc-andreessen — Unix agent architecture (LLM+shell+files+cron); death of UI; 80-year overnight success framing
  • claire-vo — 9-agent personal stack on OpenClaw; manager framing for agents; isolated machine rule

Recurring Themes

  • Skills are the product, not the model — the model is increasingly a commodity; the moat is skills, data, and domain trust
  • Design for obsolescence — write scaffolding knowing future models will need less of it; markdown > code for instructions
  • Evaluation is non-optional — generic NLP metrics don’t work for domain agents; build domain-specific test suites
  • The filesystem is a natural interface — bash + file I/O is a surprisingly powerful and general agent abstraction
  • MCP vs skills: MCP gives agents access (external systems); skills give agents judgment (how to act)
  • Agent-first software: with 100–1000x more agents than humans operating on systems, software design must optimize for agent experience — and agents select tools on actual system quality, not documentation or marketing
  • Diffusion gap: AI capability spreads faster in startups than enterprises; legacy systems (SAP, ERP) encode domain knowledge in ways that can’t be vibe-coded away

Connections

  • openclaw — open-source personal agent platform; consumer-facing instantiation of Unix agent architecture
  • data-warehouse — agents are increasingly used to query and reason over DWH data; semantic layers are a natural MCP integration point