Skip to the content.

Agent notes — agentic engineering through 2026

Rendered: https://genno-whittlery.github.io/agent-notes/

Personal study notes on agentic engineering as it evolved through 2026. The starting subject was Claude Code; the scope broadened across the year to cover adjacent agents (Hermes, opencode), local-inference backends (ds4.c / DeepSeek V4), the harness-engineering umbrella, Anthropic’s vertical surfaces (Code Review, Design, Finance), conference proceedings (AIE Europe), the economics framing (Blomfield), and the memory patterns (Karpathy’s LLM Wiki). Fifteen entries today, newest at the top; the layout is intentionally changelog-shaped so new material can be prepended without restructuring as our understanding grows.

Working discipline: every concrete claim cites a 2026 source, every hook event name is verified against the official Anthropic changelog before going in, and entries are dated so a reader (future-us included) can see which parts age first.

Entries (newest → oldest)

Entry kind: static = a dated event whose facts don’t move; rolling = a topic that accretes references and gets supersede-notes as the shape changes.

Month Kind File One-line
2026-05-11 rolling 2026-05-ds4-c-local-inference.md antirez/ds4 — single-file C inference engine running DeepSeek V4 Flash locally; OpenAI/Anthropic-compatible API so Claude Code can use it as a backend
2026-05-06 static 2026-05-code-review-and-conference.md Code w/ Claude conference; the Code Review launch
2026-05-05 static 2026-05-claude-for-finance.md Anthropic’s 10 finance-agent templates + Opus 4.7 finance-benchmark lead + Microsoft 365 integration + $1.5B enterprise JV
2026-04-17 static 2026-04-claude-design.md Claude Design + “Claude for Creative Work” — Anthropic past coding into design / Adobe-Autodesk-Ableton-Blender-Splice connector network
2026-04-08 static 2026-04-aie-europe.md AI Engineer Europe (London, 8–10 April) — harness engineering crystallises, “strange stickiness of coding products,” AI coding wars
2026-04 static 2026-04-karpathy-llm-wiki.md Karpathy’s LLM wiki Gist — spec for filesystem-first agent memory via Obsidian; spawned six+ community implementations including Cognition’s claude-memory-compiler
2026-04 rolling 2026-04-hermes-agent.md Nous Research’s Hermes Agent — CLI orchestrator that delegates to Claude Code + Codex; ACP protocol (April 2026)
2026-04 rolling 2026-04-managed-agents-sdk.md managed-agents-2026-04-01 beta — multi-instance orchestration
2026-03 rolling 2026-03-blomfield-economics.md Tom Blomfield (YC partner) on the economics of agentic coding — the “24-year-old vs Accenture” line, 95%-AI-written at YC startups, Jack Clark’s “majority of Anthropic’s code”
2026-03-31 static¹ 2026-03-source-leak-architecture.md The npm-missing-.npmignore leak — what 512K lines of CC source actually says
2026-02 rolling 2026-02-cursor-2-and-ide-agents.md Cursor 2.0 (Feb 2026) — 8 background agents on cloud VMs with git worktree isolation + Composer model; the IDE-embedded agent category (Cursor / Cline / Continue.dev / Windsurf)
2026-02 rolling 2026-02-devin-cognition.md Devin / Cognition Labs February 2026 update — parallel sessions + improved context retention; the “engineer-as-architect” thesis; canonical autonomous-end of the agent spectrum
2026 rolling 2026-agentic-engineering.md The umbrella discipline — practice of engineering reliable behaviour from non-deterministic agents; the seven sub-skills + how the other entries map to them
2026 rolling 2026-anthropic-academy.md anthropic.com/learn — first-party canonical pedagogy (Claude 101, Claude Code in action, MCP guide, certificates); the reset point when community material drifts
2026 rolling 2026-harness-engineering.md One sub-discipline of agentic engineering: designing the layer around the LLM; two-agent + three-agent patterns; the 88% production-failure stat
2026 rolling 2026-mcp-protocol.md Model Context Protocol — Anthropic-originated, donated to Linux Foundation Dec 2025; 97M+ monthly SDK downloads; the cross-vendor “USB-C for AI” connector standard
2026 rolling 2026-agents-md-standard.md AGENTS.md — cross-vendor coding-agent instructions file (Codex CLI, Copilot, Cursor, Windsurf, Amp, Devin all read it); CLAUDE.md’s sibling, also donated to Linux Foundation Dec 2025
2026 rolling 2026-evaluation-infrastructure.md Coding-agent evaluation: SWE-bench Verified, vertical benchmarks, LLM-as-judge and its biases, the 2026 production-evaluation discipline. Fills the gap flagged in the agentic-engineering umbrella entry
2026 rolling 2026-discovery-method.md Meta: how to find hot topics systematically — five-tier intake stack, the hot-topic detection rules, the 3-source verification discipline, cadence model. The methodology behind the rest of this repo
2026 rolling 2026-vibecoding-tools.md Lovable / Bolt / v0 / Replit Agent — the non-developer-audience tool category (Collins Word of the Year 2025); $100M-ARR-with-tiny-teams shape Blomfield cites
2026 rolling 2026-ai-browsers.md ChatGPT Atlas (Oct 2025) + Perplexity Comet (July 2025) — agentic-browser category; agent-drives-controlled-environment shape applied to web; Webwright is the DIY equivalent
2026 rolling 2026-aider-cli.md Aider (Paul Gauthier) — precision-edits pair programmer; 4.2× fewer tokens than Claude Code on diff-shaped tasks; TDD-with-AI discipline matured here first
2026 rolling 2026-hooks-and-effort.md Hooks evolved through 2026: effort.level, terminalSequence, duration_ms, the 8-block Stop cap
2026 rolling 2026-skills-and-marketplace.md Skills + plugin marketplace + the agentskills.io cross-vendor open standard
2026 rolling 2026-claude-md-and-plan-mode.md Forrest Chang’s CLAUDE.md (220K+ ⭐ across the fork network), plan mode, the spec-driven workflow

¹ The leak itself is a static 2026-03-31 event; the entry carries some 2026-05 supersede notes (e.g., 2.1.143 Stop 8-cap) marking post-March reverberations.

Who we follow

The intake list — bloggers, podcasts, aggregators, official release channels, adjacent ecosystems — that feeds the entries above lives in following.md. Same shape as the entries: dated, honest about gaps, cross-marked with the citations in each entry so readers can see which voices show up where.

How to read this

Top → bottom = recency-first. If you’re new to Claude Code, start at the bottom (2026-claude-md-and-plan-mode.md) — the patterns there are what everything else above is built on. If you’re already running CC daily, start at the top to catch the changes since you last paid attention.

Each entry has the same shape:

What’s not here

Updating

Add a new entry by writing <YYYY-MM>-<slug>.md and prepending one line to the index above. Don’t renumber. Don’t backdate.

When a claim in an existing entry becomes wrong (a hook event renamed, an SDK version superseded), edit in place with a strikethrough or a “superseded 2026-XX-XX” note at the top of the affected section — don’t silently delete. Half the value here is letting future-us see when the shape of Claude Code changed.

License

MIT. See LICENSE. Public so the notes are reachable across machines and shareable when useful, not because they claim to be authoritative.

About

Personal study notes kept under the Genno persona at whittlery.io. Companion to genno-whittlery/web-monkey (in-house UI fuzz harness, also kept as public notes).