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:
- Why this matters — what the story is, who’s writing about it
- The technical detail — concrete, citable claims
- Configuration or code — copy-pasteable artefacts where applicable
- Try this — one experiment you can run today to see the behaviour
- Sources — every link cited, dated when possible
What’s not here
- A from-zero install tutorial. The official docs handle that; we don’t duplicate them.
- Generic prompt-engineering advice. The 2026 consensus (and one of the entries below) is that prompt engineering has been displaced by context engineering; tips on phrasing have a short half-life.
- Anything that wasn’t documented or written about in 2026. Older patterns get a passing mention only when they’re the foundation a 2026 idea builds on.
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).