Table of Contents
Full sitemap of this reference — every chapter and major section, with deep links.
A full sitemap of this reference. Use this page when you know roughly what you want but don't remember which chapter it lives in.
For a chapter-level summary, see the Overview. For a one-week onboarding path, see Who's Who § Reading order. For continuously updated entry points to the broader field, see the Reading List.
1. Approaches
Per-system deep dives across 30+ agentic engineering products.
Agent systems
- Stripe Minions
- AgentField
- OpenHands
- Open SWE
- OhMyOpenAgent
- OpenCode
- SWE-agent
- Composio Agent Orchestrator
- Patchwork
- Goose
- Mastra
- OpenClaw + The OpenClaw Ecosystem + The Steinberger School
- Hermes Agent — deep dive
- Claude Managed Agents
- Vercel Open Agents
- OpenAI Symphony
- Rivet Sandbox Agent
- DeerFlow
- GStack
- GBrain
- Superpowers
- Everything Claude Code
- AgentHub
- Crabbox
- Clawpatch
- ClawSweeper
Cross-cutting sections inside Approaches
- Skills, Plugins & Marketplaces
- Browser-Use & Computer-Use Frameworks
- Terminal coding CLIs — the 28-CLI comparison table
2. Models
Curated model reference for agentic engineering as of May 2026.
- Decision rule before you read the tables — the 5-rule cost-discipline pattern
- Closed-source frontier — Anthropic (Opus / Sonnet / Haiku 4.x), Google (Gemini 3.x Pro / Flash / Flash-Lite), OpenAI (GPT-5.5 / 5.2 / mini), xAI (Grok 4)
- Open-weights frontier — DeepSeek V3.2 / R2, Qwen3 Max / Coder, Llama 4 Maverick / Scout, Kimi K2, GLM-5, MiniMax M2.7, Mistral Large 3
- Agent / coding specialists — GPT-5.2-codex, Devstral, Codestral 3, Qwen3-Coder, OpenCoder
- Decision shortcuts — 10-row routing table
3. Patterns
Cross-cutting architectural patterns.
- Harness Engineering — the umbrella discipline (see also the Harness Engineering deep dive)
- 1. Isolation Strategies
- 2. Orchestration Models
- 3. Context Management
- 4. Feedback Loops
- 5. Failure Recovery
- 6. Multi-Agent Coordination
4. Harness Engineering
The deep dive on what makes agents reliable.
- Why Harness Beats Model Upgrade
- The Five-Subsystem Model
- Foundations — repo-as-system-of-record · progressive disclosure · initialization · continuity
- Scope and Verification — WIP=1 · feature lists · three-layer termination · worker-vs-checker
- Observability Inside the Harness — sprint contracts · evaluator rubrics · OpenTelemetry
- The Session Lifecycle and Clean State
- The Reference Stack
- Failure-Mode Catalogue
- Decision Framework
5. Context Engineering
The named discipline of curating what's in the LLM context window.
- Why it matters — context rot and the attention budget
- The four strategies — write / select / compress / isolate
- Failure modes — poisoning · distraction · confusion · clash
- Concrete thresholds worth pinning — 95% / 85% compaction · 20K-token spill · ~12-skill ceiling
- Anti-patterns
6. Tool Design
How to write tools agents use well — starting with MCP as the assumed wire format.
- MCP — the wire format you're writing tools in — the protocol that won, runtimes (Arcade, Composio), reading list
- What "good tool design" actually means
- Consolidate, don't expose your API surface
- Compress every response — ResponseFormat enums (206 → 72 tokens)
- Lazy load — the "too many tools" problem — Tool Search Tool, -85% tokens
- Code-as-tool — give the agent a Python sandbox — 150K → 2K tokens
- SQL-over-APIs — when the right tool is a query language — Coral case study (+31% accuracy, 3.4× cost-efficient) and the cluster: Steampipe, MindsDB, CData Connect AI, PromptQL, Trino/Starburst, Cube
- Programmatic Tool Calling
- Tool Use Examples — 72% → 90%
- What to measure when iterating on tools
- Anti-patterns
7. Skills
The cross-vendor primitive for capability packaging (Anthropic open standard, Dec 2025).
- The SKILL.md format
- Progressive disclosure — the key idea
- Empirical bounds — 82% vs 9% lift · ~12-skill ceiling · 70% invocation reliability
- Designing a skill that gets invoked
- What you can ship as a skill
- Security
- Anti-patterns
8. Memory
Persistent state across turns and sessions.
- The taxonomy — three axes: lifetime / type / update mechanism
- The vendors and what they actually do — Letta · Mem0 · LangMem · LangGraph Store · Anthropic memory tool
- The filesystem-as-memory pattern
- How agents actually learn over time — three-layer continual-learning model
- Concrete patterns from production
- Anti-patterns
9. Evals
How to measure agent quality — distinct from public benchmarks.
- The mental model — three test layers: code-based / model-based / human
- How to start an eval program
- pass@k vs pass^k — the reliability gap
- Three things that silently invalidate your numbers — grading bugs · infra noise · eval awareness
- Benchmarks ≠ trustworthy by default — the ABC paper
- Categories to test
- Multi-turn eval design
- Tooling landscape — Inspect AI · LangSmith · Braintrust · Langfuse · Phoenix · Harbor
10. Benchmarks
How agentic coding is publicly evaluated.
- SWE-bench and variants — Verified, Lite, Multimodal, Multilingual, Pro
- Terminal Bench
- Inspect AI
- τ-Bench (Sierra)
- Other benchmarks worth knowing — 9-row roundup: BFCL, GAIA, BrowseComp, CORE, MLE-bench, ScienceAgentBench, OSWorld, Sweep
- Choosing a benchmark
- Benchmark-adjacent reading
11. Schools
Where does trust live? Three philosophical schools + four operational schools.
- The Central Question
- Philosophical schools:
- Side-by-Side Comparison
- Operational schools:
- Cross-Map: Operational × Philosophical
- What the Next 24 Months Look Like
12. Who's Who
29 named profiles of the people shaping the field.
- 🧠 Researchers / educators: Karpathy · Weng · Yao · Brown · Yang · Kiela · Teknium · Polosukhin
- 🔨 Operators / founders: Steinberger · Tan · Cherny · Chase · Vincent · Robinson · Liu (Beyang) · Liu (Jerry) · Schluntz · Trivedy · Martin
- ✍️ Chroniclers / synthesizers: Willison · Osmani · Mollick · swyx + Fanelli · Husain · Yan · Huyen · Schmid · Wolfe · Schulhoff
- Appendix — additional candidates
- Reading order if you're new — one-week onboarding path
13. Organizations
How companies organize around agents.
- The Stripe Model
- The Open-Source / Startup Model
- Organizational Patterns
- The Infrastructure You Need
- The Future
14. Inference
LLM inference solutions.
- Direct API Providers
- Inference Platforms
- Nebius AI Cloud — deep dive
- Routing & Gateway Solutions
- Self-Hosted Inference
- tinygrad / the tiny corp — deep dive
- Inference Strategy for Agents
- Decision Framework
15. Sandboxes
The execution-environment layer.
- Why Sandboxes Matter for Agents
- The Sandbox Market Structure — four-layer model
- Core Use Cases
- Isolation Tiers
- Purpose-Built Agent Sandboxes — full vendor table
- Contree — The Git-Native Sandbox — deep dive
- Cloud Development Environments (CDEs)
- Open-Source Isolation Primitives
- Agent Patterns Enabled by Modern Sandboxes
- Decision Framework
- Integration Examples
16. Hosting & Execution Infrastructure
150+ vendors across 9 major categories.
- Agent Hosting & Execution Platforms — the six-tier decision framework
- Code Execution Sandboxes — quick-ref table
- Turnkey Managed Platforms — OpenClaw-native, enterprise hubs, no-code builders, Autonomous Coding Agents, visual IDEs
- Agent-Optimized Hosting
- Agent Orchestration — durable execution, cloud workflows, agent-specific frameworks, data/ML orchestrators
- Cloud Mac Hosting
- CI Runners for Agent Iteration — Blacksmith, Tenki Runners, Depot, Namespace, BuildJet, RunsOn, Ubicloud
- Self-Hosted Infrastructure — GPU clouds, general clouds, VPS for agents
- Agent Memory & Context Infrastructure — purpose-built memory, vector DBs, graph DBs
- Agent Observability & Evaluation — tracing, evaluation, guardrails
- MCP Servers, Registries & Gateways
- Agent Identity, Auth & Secrets
- Choosing Your Stack — starter / growth / scale / enterprise
- Decision Framework
16.5 Tenki Review
Vendor deep-dive on Tenki — the only operator shipping all three legs of the agent CI loop (Sandbox, Runners, Code Reviewer) — benchmarked head-to-head against the category leader for each leg.
- Company — Luxor origin, team, funding, customers, security
- The agent-iteration loop — the bundle thesis
- Sandbox — vs E2B, Daytona, Modal, Blaxel, Sprites.dev, Contree
- Runners — vs Blacksmith, Depot, Namespace, BuildJet, RunsOn, Ubicloud
- Code Reviewer — vs CodeRabbit, Greptile, Qodo, Copilot review, Cursor BugBot, Graphite Agent
- Strategic position — the bundle thesis examined honestly
- Where DevRel adds the most leverage
- Verdict
17. Generative UI
The agent's front-end story.
- Why it matters for agentic engineering
- The three primary patterns — Static, Declarative, Open-ended
- Specifications and protocols — A2UI, AG-UI, MCP-UI, Open-JSON-UI
- Frameworks — CopilotKit (the reference example), Vercel AI SDK, Mastra + CopilotKit
- Code examples — static, declarative, open-ended
- Trade-offs — consistency vs. flexibility
- Decision framework
- A2UI adoption snapshot
18. Research Notes
Source-of-truth bibliography behind every page above. 100+ primary sources ingested in May 2026; structured per-URL digest with key claims, frameworks named, and which slot in the reference each source fills.
- Anthropic Engineering (19 URLs)
- LangChain Blog (20 URLs)
- Individual articles + arxiv + courses (10 URLs)
- GitHub repos + framework docs (21 URLs)
- People's blogs + newsletters + podcasts (14 URLs)
- Tools, platforms, courses, communities (24 URLs)
- Cross-cutting findings — 7 patterns that repeated across enough sources to pin
Meta pages
These don't fit the numbered chapter sequence but are linked from the sidebar Get Started group:
- Reading List — curated entry points to follow the field (newsletters, blogs, podcasts, courses, communities, conferences, reference repos), with a practical weekly cadence at the bottom
- Changelog — what's been added to this site, newest first; content additions only (bug fixes / refactors / UX live in git log)
Cross-page indexes
- Schools framing: introduced in Approaches § The Steinberger School, formalized in Schools, referenced from Who's Who profiles
- Context engineering thread: Context Engineering coins the discipline; Tool Design is the action-layer slice; Skills is the capability-packaging primitive; Memory is the durable-state layer
- Evaluation thread: Evals covers the methodology (your tests against your failure modes); Benchmarks covers the public leaderboards (SWE-bench, Terminal Bench, etc.)
- Vendor cross-reference: many vendors appear in both Sandboxes and Hosting & Execution — the Sandboxes page is the deep dive, Hosting & Execution is the quick reference
- Reading order for newcomers: Who's Who § Reading order for the one-week onboarding path; Reading List for the broader source map