Changelog

What's been added to this site, newest first. Bug fixes, refactors, accessibility passes, sidebar restructures, and star-count refreshes happen in git history but aren't tracked here — this page is for content additions so you can see what's new to read.

Each entry links to the PR that shipped it. Authors: append your entry to the top under the current date when you merge.

This page has two timelines:

1. Site additions — when content was added to this site (newest first), keyed by PR.
2. Primary source timeline — when the underlying primary sources were originally published (oldest first), keyed by date.

---

Primary source timeline

Publication dates of the primary sources that informed the polished pages. Compiled from Research Notes (re-verify dates at primary source before citing). Most-impactful entries marked with ★ — these are the canonical-reference posts the rest of the site builds on.

2024

• 2024-06-06 — ★ The Prompt Report: A Systematic Survey of Prompting Techniques The field's reference survey — 1,500+ papers, 200+ techniques, co-authored with OpenAI/Microsoft/Google/Princeton/Stanford; the canonical citation when claiming "X prompting trick works"
• 2024-12-19 — ★ Building Effective Agents Foundational pattern taxonomy — the entry point article for the whole site
• 2024-12-20 — Building effective agents (Simon Willison's commentary) The de facto vocabulary the field now uses for workflow-vs-agent design — required reference taxonomy

2025

• 2025-01-14 — Introducing Ambient Agents UX/architecture pattern — provides the canonical "non-chat" agent interaction model and HITL vocabulary
• 2025-04-20 — How to Think About Agent Frameworks Framework taxonomy — clarifies "orchestration vs abstraction" axis when picking tools
• 2025-06-13 — ★ How We Built Our Multi-Agent Research System The canonical orchestrator-worker case study with token-economics numbers
• 2025-06-14 — How we built our multi-agent research system (Simon's commentary) Single best public case study on production multi-agent research — concrete numbers (15× tokens, 90.2% lift, 90% time cut)
• 2025-06-23 — Inspect AI, An OSS Python Library For LLM Evals Concrete recommendation for the eval layer of an agent stack — production-grade, framework-agnostic
• 2025-07-02 — ★ Context Engineering for Agents Foundational vocabulary — the write/select/compress/isolate taxonomy is the canonical framing
• 2025-07-03 — ★ Establishing Best Practices for Building Rigorous Agentic Benchmarks (ABC paper) Provides a vetted checklist for trustworthy agent evals — addresses the "looks great on benchmark, fails in prod" gap
• 2025-07-09 — How to Build an Agent Onboarding — canonical "how do I start" sequence; useful as a default reading-list entry
• 2025-09-11 — ★ Writing Effective Tools for Agents — with Agents Reference for the tool-design chapter — "tools are software for agents" with measurement framework
• 2025-09-29 — Building Agents with the Claude Agent SDK Canonical "what is an agent loop" reference — slot directly under harness fundamentals
• 2025-09-29 — ★ Effective Context Engineering for AI Agents Founding doc for the context-engineering chapter — coins the "attention budget" framing
• 2025-10-16 — ★ Equipping Agents with Agent Skills Canonical reference for the Skills mechanism — paired with code-execution-with-mcp
• 2025-10-20 — Beyond Permission Prompts: Sandboxing Claude Code Reference architecture for the security/isolation slot — paired with auto-mode for the autonomy story
• 2025-10-28 — Doubling Down on Deep Agents Origin of the "harness" framing — establishes Deep Agents as a harness, not a framework competitor
• 2025-11-04 — ★ Code Execution with MCP Argument and numbers for replacing direct tool calls with code-as-tool — central to context-engineering chapter
• 2025-11-24 — Advanced Tool Use on the Claude Developer Platform Concrete numbers for the "tools as context burden" pattern — supports the case for lazy tool loading
• 2025-11-26 — Effective Harnesses for Long-Running Agents Concrete recipe for cross-session continuity in long-running coding agents
• 2025-12-02 — State of Agent Engineering (LangChain survey, fielded Nov 18 – Dec 2) Hard 2025 industry baseline — production %, observability %, model-mix, and use-case mix you can quote
• 2025-12-03 — Evaluating Deep Agents — Our Learnings Eval lesson — concrete proof that single-step decision evals are the high-leverage entry point

2026

• 2026-01-09 — Demystifying Evals for AI Agents Canonical eval-design primer — paired with the BrowseComp and infrastructure-noise posts
• 2026-01-21 — Designing AI-Resistant Technical Evaluations Eval-design lesson — how to write tasks that still discriminate humans from frontier models
• 2026-01-28 — Context Management for Deep Agents Deep-agents lesson — concrete numbers (20K tokens, 85% threshold, 10-line preview) for compress + isolate
• 2026-02-05 — Building a C Compiler with a Team of Parallel Claudes Reference example of multi-agent autonomous large-codebase work with cost/scale numbers
• 2026-02-05 — Quantifying Infrastructure Noise in Agentic Coding Evals Critical eval-hygiene caveat — required reading before citing leaderboard deltas
• 2026-02-12 — On Agent Frameworks and Agent Observability Observability framing — the "logic lives in traces" line is the strongest one-liner for why observability is first-class
• 2026-02-17 — Improving Deep Agents with Harness Engineering Harness pattern — flagship case study showing harness changes alone can deliver double-digit benchmark gains
• 2026-03-05 — Evaluating Skills Skills design — the ~12-skill ceiling and 70% invocation reliability are useful empirical bounds
• 2026-03-06 — Eval Awareness in Claude Opus 4.6's BrowseComp Performance Eye-opening eval-integrity case study — model behavior + how benchmarks leak
• 2026-03-10 — ★ The Anatomy of an Agent Harness Definitional — the canonical "what is a harness" reference; sets the vocabulary the other harness-engineering posts build on
• 2026-03-24 — Harness Design for Long-Running Application Development Reference for multi-agent harness design with real cost/time numbers — pairs with effective-harnesses post
• 2026-03-25 — Claude Code Auto Mode: a Safer Way to Skip Permissions Concrete pattern for replacing permission prompts with classifier-mediated autonomy
• 2026-03-26 — How Middleware Lets You Customize Your Agent Harness Harness pattern — the six-hook taxonomy is the canonical reference for "where do I put this customization?"
• 2026-03-26 — How We Build Evals for Deep Agents Eval architecture — concrete category and metric scheme to copy when designing a project's eval taxonomy
• 2026-03-27 — Agent Evaluation Readiness Checklist Eval lesson — readiness gate before automating eval infra; pairs with the "harness hill-climbing" post
• 2026-04-02 — ★ Open Models Have Crossed a Threshold Model behavior — concrete cost/perf table for the "should I use open models?" decision
• 2026-04-03 — Production Agents Self-Heal Ops pattern — concrete recipe for an auto-remediation loop using a coding agent as the fixer
• 2026-04-05 — Continual Learning for AI Agents Mental model — the three-layer separation is a clean way to organize "how agents improve" sections
• 2026-04-07 — Deep Agents v0.5 Version milestone — async-subagent primitive is the v0.5-specific feature worth pinning
• 2026-04-08 — Better Harness — Hill-Climbing with Evals Harness pattern — concrete recipe for moving harness changes through eval gates rather than vibes
• 2026-04-08 — Scaling Managed Agents: Decoupling the Brain from the Hands Defines a reference architecture (session/harness/sandbox) for production agent systems
• 2026-04-16 — The Complete Guide to Harness Engineering (claudecode-lab) Most concrete public reverse-engineering of Claude Code's harness — transferable blueprint
• 2026-05-25 — Andrej Karpathy joins Anthropic Industry-signal post: the person who coined "vibe coding" and "agentic engineering" returns to R&D inside the lab that publishes most of the primary sources on this page

Total: 45 dated primary sources. See Research Notes for the full structured digest per source (key claims, frameworks named, quotable lines, frameworks/concepts cataloged).

Sources not in this timeline: vendor / framework docs pages (no single publish date — continuously updated); GitHub repo READMEs (use last-commit date if needed); blog/newsletter homepages (use most-recent-post date); YouTube channels and podcast index pages. All catalogued in Research Notes regardless.

---

Site additions

2026-06-20

New "Tenki Review" vendor deep-dive page

New page at tenki.md, wired into the Infrastructure sidebar group as Tenki Review. Vendor-named pages are unusual on this site (Nebius and Coral got dedicated callouts inside category pages, but no standalone vendor page exists for any other vendor), and this one is justified by Tenki being the only operator shipping three of the four legs of the agent CI loop — Sandbox, Runners, and Code Reviewer — under one roof. The page benchmarks each product head-to-head against its category leader: Sandbox vs E2B / Daytona / Modal / Blaxel / Sprites.dev / Contree; Runners vs Blacksmith / Depot / Namespace / BuildJet / RunsOn / Ubicloud; Code Reviewer vs CodeRabbit (the mindshare leader, $60M Series B Sep 2025, ~$40M ARR, 8K paying customers) / Greptile ($25M A from Benchmark) / Qodo (OSS / self-hostable) / GitHub Copilot review (first-party threat) / Cursor BugBot / Graphite Agent / Sourcery / Ellipsis / Bito. Closes with the strategic-position section examining the bundle thesis (where it wins: vendor-stitching pain; where it loses: best-of-breed depth) and a "where DevRel adds the most leverage" section that's analytically framed but pulls no punches on the gaps (no GPU sandbox, no public Runners benchmark, undisclosed reviewer models, GitHub-only across the line, no marquee enterprise reference). Cross-links to Sandboxes, CI Runners for Agent Iteration, and the Nebius / Contree comparable deep-dives.

2026-06-13

New "CI Runners for Agent Iteration" section in Infrastructure

New top-level section in Infrastructure § CI Runners for Agent Iteration placed between Cloud Mac Hosting and Self-Hosted Infrastructure. Opens with the framing that CI is the agent's feedback loop — when an agent fleet opens 30 PRs an hour and each waits 8 minutes on a GitHub-hosted runner, the harness is throttled by the runner queue, not the model. Covers seven drop-in runs-on: replacements: Blacksmith (YC W24, $10M Series A Sep 2025 GV-led, "Unblock AI Development with Fast CI" — used by Vercel / Supabase / Clerk / Mercury / Ashby), Tenki Runners (Luxor / Seattle, Firecracker microVM, same operator as the Tenki Sandbox for agent code execution), Depot (Docker / Bazel cache-first), Namespace (runners + remote dev envs), BuildJet (the mature conservative pick), RunsOn (self-hosted on your AWS), and Ubicloud (OSS, "open AWS"). Includes a choosing rubric and cross-references to Mac CI Runners and the Sandboxes branching-as-replay-buffer pattern. Fills a previously missing leg of the agent-iteration stack — the site already covered sandboxes, dev environments, and orchestration but not CI itself as an agent infrastructure category.

tinygrad / the tiny corp added to Inference

New row in Inference § Self-Hosted Inference for tinygrad alongside a dedicated § tinygrad / the tiny corp — Standout Local-AI Stack callout (mirroring the existing Nebius pattern). Covers the 20K-LOC framework with its unified UOp IR and pattern-rewrite kernel fusion, the full backend matrix (CUDA / AMD / NV / Metal / WebGPU / Qualcomm DSP / Clang / LLVM, plus the NVK / Mesa NIR path from v0.12), the tinybox hardware line (red v2, green v2, pro, exabox), the AMD-sovereignty angle (the public 7900 XTX firmware fight, Lisa Su's intervention, the "completely sovereign compute stack" Phoronix reported on), production usage at comma.ai openpilot (Snapdragon 845 on-device + tinybox-trained world model), and status (33K stars, v0.13.0 May 2026, still pre-1.0, $5.1M raised, San Diego, run by George Hotz). The site previously had no entry for the framework + hardware + sovereignty axis — this fills it without sprawling into a new chapter.

2026-05-26

Merge.dev + enterprise SaaS-integration platforms added to Infrastructure / Inference

New "Enterprise SaaS-integration platforms (now agent-capable)" subsection in Infrastructure § MCP covering Merge, Nango, Paragon, Apideck, and Scalekit — companies that started in unified-API / embedded-iPaaS and pivoted into agent tool-calling once MCP won. Closing paragraph names the strategic convergence (all racing toward "integration + tool-calling layer for enterprise agents" from different starting points) and the architectural critique competitors make of Merge specifically (auth model built for developer-initiated API calls; agents need per-user delegation + sub-second scoped execution). Merge's LLM Gateway product also added as a new row in Inference § Routing & Gateway Solutions. Coverage gap surfaced via a Merge-DevRel competitive-analysis dump the user shared — picked the 5 named not-covered competitors; intentionally skipped Toolhouse / Finch / Kombo / StackOne / Workato (long tail) and the Unified API category itself (borderline for the site's scope).

Coral: Infrastructure entry + Tool Design "Risks to watch" callout

Follow-up to the SQL-over-APIs section from PR #29. Added Coral as a named product row in Infrastructure § MCP Servers, Registries & Gateways (where MCP-native data products belong); the row notes the architectural difference from the rest of the table (SQL query vs tool-per-action) and flags integration breadth as the load-bearing risk. Also added a "Risks to watch if you're betting on this pattern" callout to the existing Tool Design SQL-over-APIs section, distilling four load-bearing concerns from the broader SWOT material: integration deficit, token-efficiency time-window, MCP single-vendor protocol risk, and local-first vs centralized-governance trade-off. Intentionally NOT added: SPAs, Magic Quadrant positioning, pairwise SWOTs, or 90/180/365-day strategic recommendations — those don't fit the site's pattern-first reference character.

Tool Design: new "SQL-over-APIs" section comparing Coral and the cluster

Added a new H2 section in Tool Design for the architectural pattern that extends code-as-tool: expose a single SQL endpoint instead of N micro-tools per data source. Coral (withcoral.com) is the case study — Apache 2.0, Rust, local-first, MCP-native, with a published benchmark (+31% accuracy, 3.4× more cost-efficient, −42% latency vs direct provider MCPs across 82 real Claude Code tasks). Comparison table covers the cluster: Steampipe (AGPL, 150+ plugins), MindsDB (Python, ML-baked), CData Connect AI (proprietary, 300+ drivers), PromptQL (Hasura, agent-outcome layer), Trino / Starburst (lakehouse federation), Cube (semantic layer). Includes "when SQL-over-APIs is right" and "when tool-per-call is still right" decision rules. Cross-links to Memory § Context Hub (the adjacent curated-knowledge pattern) and the existing Infrastructure MCP section.

docs/ directory removed

The parallel Starlight setup at docs/ (which was never deployed — Vercel serves the root) is gone. 34 tracked files removed. All unique content was preserved into content/ first via PR #26: the 3 comparison tables that commit 1bb1bcb had accidentally deleted from approaches.md were restored, and the 24 SEO descriptions from .mdx frontmatter were ported as HTML comments at the top of each content/*.md. Build pipeline (build.sh + index.html) is unchanged and unaffected.

Context Hub + Moltbook added across Memory / Context Engineering / Infrastructure / Research Notes

Two additions surfaced via The Batch newsletter:

• Context Hub (Andrew Ng, 13.4K stars, MIT) — agent-readable curated API docs with a chub CLI for search / fetch / annotate / upvote. Added to Memory § Vendors as a knowledge-layer entry and Context Engineering § Write as the "externalized knowledge registries" pattern.
• Moltbook — "the front page of the agent internet": agent social network + identity provider. Added to Infrastructure § Identity/Auth as a speculative consumer-shaped agent-identity entry.

Both added to Research Notes under a new Section 7 (post-ingestion additions).

2026-05-25 (later)

Who's Who: Sander Schulhoff profile added — #26

Sander Schulhoff — the OG prompt engineer. Created LearnPrompting.org two months before ChatGPT's release (now 3M+ users); co-authored The Prompt Report with OpenAI / Microsoft / Google / Princeton / Stanford / CMU researchers (1,500+ papers, 200+ techniques surveyed). Also ran HackAPrompt (EMNLP 2023 Best Theme Paper on prompt injection). Now CEO of InventoryQuant (YC W26). Profile added under the chronicler+researcher hybrid archetype. The Prompt Report also added to the Primary Source Timeline as a 2024-06-06 ★ entry. Total profiles: 28 → 29. Total dated sources: 44 → 45.

URL canonicalization — #26

Migrated stale URLs to current canonical forms across all content:

• LangChain blog: blog.langchain.com/<slug>/ → www.langchain.com/blog/<slug> (70 occurrences across 6+ files)
• LangChain self-heal post: slug renamed production-agents-self-heal → how-my-agents-self-heal-in-production (same post, same Apr 3 2026 date — confirmed via fetch)
• ECC repo: github.com/affaan-m/everything-claude-code → github.com/affaan-m/ECC (repo renamed; section anchor #everything-claude-code preserved to avoid breaking internal links — rename noted in approaches.md body as "Also known as: ECC")

One blog.langchain.com reference intentionally left in research-notes.md — it documents the redirect explicitly.

Who's Who: Cameron R. Wolfe profile added — #26

Cameron R. Wolfe (PhD, ML) was in reading-list.md and research-notes.md but missing from Who's Who. His Deep (Learning) Focus Substack (68K+ subscribers) is the research-paper-translator voice the site was missing — the canonical explainer source for MoE architectures, reasoning LLMs, test-time-compute scaling, and similar concepts cited across Models and elsewhere. Total Who's Who profiles: 27 → 28.

Who's Who: 2 new profiles (Huyen, Schmid) — #26

Coverage audit surfaced two real gaps in Who's Who: Chip Huyen was mentioned only in the appendix ("considered but didn't profile") despite AI Engineering being the standard production-ML reference; Philipp Schmid (Google DeepMind, formerly Hugging Face) was missing entirely despite being one of the highest-cadence recipe-author voices in the field. Both promoted to full chronicler profiles with key works + "start here" entries. Total profiles: 25 → 27.

Who's Who: Karpathy joins Anthropic — #26

Updated Andrej Karpathy's career line to reflect his May 2026 announcement that he's joined Anthropic to "get back to R&D" at the frontier of LLMs. Eureka Labs era ends; Zero-to-Hero lecture series continues as his ongoing education work. Notable industry signal — the person who coined both "vibe coding" and "agentic engineering" is now inside the lab that publishes most of the primary sources cited across this reference.

Tool Design: MCP elevated to lead section — #26

MCP was buried at section 9 of 10. Promoted to the first major section after the intro ("MCP — the wire format you're writing tools in") and substantially expanded with: the Nov 2024 Anthropic launch + late-2025 adoption inflection, the "you're writing MCP whether you know it or not" framing, an anatomy-of-an-MCP-tool example, the Arcade / Composio governance runtime layer, and a curated reading list. Table of Contents reordered to match.

Changelog: added Primary Source Timeline — #26

43 dated primary sources from research-notes.md arranged chronologically (Dec 2024 → Apr 2026), each with a one-line topic note and the canonical reference posts marked with ★. The page now has two timelines: site additions (when we added it) and primary sources (when the field originally published it).

Lost-content rescue + docs/ preservation — #26

Audit of docs/ vs content/ found that commit 1bb1bcb had accidentally deleted three tables (Feature Matrix, Capability Breakdown, Composability) from approaches.md that PR #3 explicitly merged in. Only surviving copy was in docs/comparison.mdx (the never-deployed Starlight parallel). Restored to end of approaches.md. Also ported all 24 SEO description fields from docs/.../*.mdx frontmatter into content/*.md as invisible HTML comments so the metadata survives any future migration.

2026-05-25

Six research-driven Landscape pages — #24

Distilled from the 109-source ingestion in #23. New pages: Context Engineering, Tool Design, Skills, Memory, Evals (Landscape group); Reading List (Get Started group). Each page anchors to specific empirical numbers — e.g. Tool Search Tool's -85% token reduction, Skills' 82% vs 9% lift, the pass@k vs pass^k reliability gap.

Research Notes bibliography — #23

Structured digest of 109 primary sources across six parallel ingestion passes (Anthropic Engineering, LangChain blog, individual articles + arxiv, GitHub repos + docs, blogs + newsletters, tools + platforms). Per-entry format: URL, fetch status, key claims with specific numbers, frameworks named, "which slot it fills." Lives under the new Reference sidebar group.

Models reference page — #22

New Models page (Landscape group) with three tables: closed-source frontier, open-weights frontier, agent/coding specialists. Anthropic and Google prices verified at primary source; OpenAI/xAI/OSS rows link to lab pricing pages. Includes a 5-rule cost-discipline decision rule and a "decision shortcuts" routing table.

Who's Who: 5 new profiles — #21

Added Hamel Husain (eval methodology), Eugene Yan (applied LLM systems), Erik Schluntz (Claude Code), Sahil Trivedy (Operator), Charlie Martin (CodeBuff). Brings Who's Who from 20 to 25 profiles.

Benchmarks: 9-row "Other benchmarks worth knowing" roundup — #20

Added BFCL, GAIA, BrowseComp, CORE, MLE-bench, ScienceAgentBench, OSWorld, Sweep, plus a caveats section on infrastructure noise and eval awareness.

Five framework entries to Approaches — #19

Added Claude Agent SDK, Deep Agents, DSPy + GEPA, Smolagents, and the LangChain create_agent primitive. Closes the gap-pass II checklist from the roadmap.

Harness Engineering: First Principles preamble — #17

Five-claim preamble at the top of harness-engineering.md anchoring the rest of the page (model ≠ system, harnesses encode stale assumptions, etc.).

Events page + Events/GitHub icon buttons — #18

New Events page (Get Started group) with the agent-event calendar.

Roadmap gap-pass — #16

Added Letta, Inngest + Temporal, Tavily, Inspect AI, τ-bench, plus 5 additional frameworks/agents to fill earlier roadmap gaps.

---

2026-05-24 (and earlier May 2026)

Chrome DevTools MCP added to Browser-Use frameworks — #13

New row in the Browser-Use & Computer-Use Frameworks table on Approaches.

Bumblebee (Perplexity) and Kimi Agent Swarm (Moonshot) added to Infrastructure — #11

Two new entries in the Infrastructure chapter.

PostHog Code added to Approaches — #10

New commercial / proprietary Coding CLI entry.

Generative UI chapter

New chapter under Interfaces, with CopilotKit + A2UI as the reference example.

Nebius events + SIGGRAPH 2026 added to Events

Extends the event calendar with sponsored and conference entries.

Events page (initial)

First version of the Events page with global map + chronological list.

Table of Contents page

Top-level TOC under Get Started, with a "this reference" framing sweep.

Who's Who expansion 12 → 20 + Schools page added

New Schools page (Landscape group) for intellectual lineages. Who's Who grew from 12 to 20 named profiles.

Who's Who page (initial)

First version with 12 profiles + suggested reading order.

Hermes Agent deep dive

Long-form treatment of Hermes Agent inside the harness landscape.

Superpowers, Everything-Claude-Code, Hermes, DeerFlow + Steinberger School

Five new agent / harness entries plus the Steinberger School treatment.

GStack, GBrain, AgentHub

Three end-to-end Claude Code harnesses added to Approaches.

Harness Engineering deep dive page

First version of the Harness Engineering page (Landscape group).

Warp Oz + Warp ADE

New Warp entries in Approaches.

---

2026-05-07

Four new chapters: Skills/Plugins, Browser-Use, MCP, Identity/Auth

Foundational chapters added across Approaches and Infrastructure.

Q4'25-Q2'26 entries to existing categories

Backfill of recent releases across multiple chapters.

Blaxel, Orchestrator.build, SmolVM, Kubernetes Sandbox CRD

Four new Infrastructure entries.

---

2026-04-20

VPS-for-agents deep dive

New Hosting & Execution treatment.

---

How to update this page

When you merge a content-addition PR, prepend an entry at the top under the current date:

## YYYY-MM-DD

### Short title — [#NN](link-to-PR)
1-2 line description of what was added. Anchor to a specific number,
framework, page, or section so the entry is scannable.

Skip entries for:

• Bug fixes (broken links, typos, build errors)
• Pure refactors (page merges, sidebar reorders without new content)
• UX / accessibility / mobile / styling
• Star count refreshes
• Tooling (CI, lint, build config)

These all live in git log if anyone needs them.