Operations Log#

Chronological record of wiki operations. Each entry uses the format: ## [YYYY-MM-DD] operation | Subject

Operations: ingest, query, lint, compile, maintenance

[2026-04-10] ingest | LLM Knowledge Bases#

Raw: LLM Knowledge Bases
Actions: Created concept LLM-as-Compiler Knowledge Base, added to index

[2026-04-10] ingest | LLM Wiki (Karpathy's Gist)#

Raw: llm wiki
Actions: Updated concept LLM-as-Compiler Knowledge Base, added to index

[2026-04-10] ingest | Best Practices for Claude Code#

Raw: Best Practices for Claude Code
Actions: Created concept Claude Code Best Practices, added to index

[2026-04-10] ingest | Harness Engineering (OpenAI)#

Raw: Harness engineering leveraging Codex in an agent first world
Actions: Created concept Agent Harness Engineering, added to index

[2026-04-10] ingest | Effective Harnesses for Long-Running Agents#

Raw: Effective harnesses for long running agents
Actions: Updated concept Agent Harness Engineering, added to index

[2026-04-10] ingest | Claude Mythos Preview#

Raw: Claude Mythos Preview / red anthropic com
Actions: Created concept LLM-Driven Vulnerability Research, added to index

[2026-04-10] query | What are AI tools?#

Output: What Are AI Tools?
Actions: Filed answer in derived/

[2026-04-10] maintenance | Setup restructure#

Actions: Consolidated per-topic indexes into single the index, added log.md, expanded compiler-prompt and workflow to align with llm-wiki pattern

[2026-04-10] lint | Full wiki audit#

Actions: Added missing backlink LLM-Driven Vulnerability Research → LLM-as-Compiler Knowledge Base (reverse direction). No broken links, no orphan pages, no index drift, no thin articles. Flagged missing page candidates and data gaps for future compilation.

[2026-04-14] lint | Full wiki audit#

Actions: Clean run — no broken links, orphans, missing backlinks, index drift, or thin articles. Updated index statistics (raw count 6→8) and added Pending Ingest section listing two uncompiled raw documents (AgentOpt v0.1, Brevity Constraints). Bumped updated and last-lint frontmatter to 2026-04-14.
Flagged (no fix): Entity page candidates (Andrej Karpathy, Anthropic, OpenAI Codex, Project Glasswing, qmd CLI, MCP, Obsidian Web Clipper, CLAUDE.md/AGENTS.md schema) — defer until second source mentions each.

[2026-04-14] ingest | AgentOpt v0.1 Technical Report#

Raw: AgentOpt v0 1 Technical Report Client Side Optimization for LLM Based Agent
Actions: Created concept Client-Side Agent Optimization; added bidirectional links from Agent Harness Engineering, Claude Code Best Practices, LLM-Driven Vulnerability Research, LLM-as-Compiler Knowledge Base; added index entries under Concepts and Source Documents.

[2026-04-14] query | When is best to use Opus 4.6 for work?#

Output: When to Use Claude Opus 4.6 for Work
Actions: Filed answer in derived/ synthesizing Client-Side Agent Optimization, Scale-Dependent Prompt Sensitivity, and Claude Code Best Practices; added Derived table entry and bumped statistics (1→2 derived outputs). Adding backlinks from the three concept articles.

[2026-04-14] ingest | Brevity Constraints Reverse Performance Hierarchies#

Raw: Brevity Constraints Reverse Performance Hierarchies in Language Models
Actions: Created concept Scale-Dependent Prompt Sensitivity; added bidirectional links from Agent Harness Engineering, Claude Code Best Practices, LLM-Driven Vulnerability Research, and cross-reference with Client-Side Agent Optimization (same phenomenon from opposite framings); added index entries.

[2026-04-17] ingest | Auto mode for Claude Code#

Raw: Auto mode for Claude Code
Actions: Created concept Claude Code Auto Mode; expanded existing auto-mode mentions in Claude Code Best Practices into backlinks; added bidirectional connection to Agent Harness Engineering (mechanical-invariant enforcement) and LLM-Driven Vulnerability Research (secondary-model classifier pattern); added index entry under Concepts and Source Documents.

[2026-04-17] ingest | Introducing Claude Opus 4.7#

Raw: Introducing Claude Opus 4 7
Actions: Created entity Claude Opus 4.7; updated Claude Code Best Practices (tokenizer inflation + xhigh default amplify context-budget constraint); updated LLM-Driven Vulnerability Research with Project Glasswing update section naming Opus 4.7 as first post-Glasswing GA model; added 4.7 addendum to When to Use Claude Opus 4.6 for Work reframing five decision rules as hypotheses-to-retest rather than invalidating them. Cross-linked with Claude Code Auto Mode (auto mode extended to Max users at 4.7 launch). Added index entries under Concepts and Source Documents; bumped statistics (raw 8→10, concepts 6→8).

[2026-04-17] query | Opus 4.6 vs 4.7 deltas and multi-agent coding considerations#

Output: Opus 4.6 → 4.7 Changes and Multi-Agent Coding Considerations
Actions: Filed answer synthesizing Claude Opus 4.7, Claude Code Best Practices, Claude Code Auto Mode, Client-Side Agent Optimization, Scale-Dependent Prompt Sensitivity, Agent Harness Engineering, and When to Use Claude Opus 4.6 for Work addendum. Added Derived backlinks in all six contributing concept articles. Added Derived table entry and bumped statistics (derived 2→3).

[2026-04-28] ingest | Symphony — open-source Codex orchestration spec (OpenAI)#

Raw: An open source spec for Codex orchestration Symphony
Actions: Created entity Symphony and concepts Ticket-Driven Agent Orchestration and Codex App Server Protocol. Updated Agent Harness Engineering with new "Harness as Service" section covering daemon-first deployment, per-tenant workspace isolation, container-as-trust-boundary, and Symphony's "objectives, not transitions" lesson as the orchestration-layer restatement of "enforce invariants, not implementations." Updated Claude Code Best Practices with cross-tool capability mapping (Claude Code / Hermes / Codex) and the session-vs-daemon deployment-architecture distinction. Updated LLM-as-Compiler Knowledge Base with new "Spec-as-Compilation Source" section covering Symphony's cross-language spec-fuzzing technique (compile SPEC.md in 6 languages, use divergences to find ambiguities). Added bidirectional Connections in all four updated articles. Added 4 index Concepts entries (symphony, ticket-driven-agent-orchestration, codex-app-server-protocol, hermes-agent — the last anticipating the next ingest) and 1 Source Documents entry. Statistics: raw 10→11 mid-pass.

[2026-04-28] ingest | Hermes Agent — Tips & Best Practices (Nous Research)#

Raw: Tips & Best Practices
Actions: Created entity Hermes Agent (jointly with the Telegram tutorial ingest below) covering CLI surface, AGENTS.md/SOUL.md context-file split, bounded MEMORY.md/USER.md memory model, frozen-snapshot semantics, /compress and delegate_task levers, prompt-cache discipline, container-backend safety model. Updated Client-Side Agent Optimization with new "Manual Levers in the Wild" section mapping Hermes commands (/model, /compress, delegate_task, bounded memory, prompt-cache discipline) to AgentOpt's combo-space levers — they're the same levers exposed manually. Added Connections from Claude Code Auto Mode (different approval-model design) and from each updated article. Added 1 Source Documents entry.

[2026-04-28] ingest | Hermes Agent — Team Telegram Assistant Tutorial (Nous Research)#

Raw: Tutorial Team Telegram Assistant
Actions: Extended entity Hermes Agent with Gateway daemon details: systemd/launchd installation, allowlist vs. DM-pairing authorization (codes, expiry, rate limiting, lockout), home channel + cron job pattern, fresh-session caveat for cron prompts, container backend ($5/month VPS sizing, macOS launchd PATH gotcha). Reinforced Connections to Symphony (parallel always-on daemon, per-issue vs. per-user tenancy) and Agent Harness Engineering (Hermes Gateway is canonical "harness as service" example alongside Symphony). Added 1 Source Documents entry. Final statistics: raw 10→13, concepts 8→12 (added symphony, ticket-driven-agent-orchestration, codex-app-server-protocol, hermes-agent).

[2026-04-28] compile | Cross-link sweep#

Actions: Scanned remaining concept articles (Claude Code Auto Mode, LLM-Driven Vulnerability Research, Scale-Dependent Prompt Sensitivity, Claude Opus 4.7) for natural backlinks to the four new pages. Added one forward link from Claude Code Auto Mode to Hermes Agent (different approval-model design point). Skipped the others — connections too tenuous, would have been link-stuffing rather than compounding knowledge. Deferred Agent Context Files concept page to a separate compile pass: the cross-vendor markdown-as-control-plane pattern (CLAUDE.md / AGENTS.md / SOUL.md / WORKFLOW.md / SPEC.md / .cursorrules) deserves a dedicated sweep across 6+ existing articles rather than being bundled here.

[2026-05-06] ingest | Boris Cherny at Sequoia AI Ascent 2026 (YouTube)#

Raw: anthropics boris cherny why coding is solved
Actions: Co-compiled with two other 2026 sources (see entries below). New entity Boris Cherny. Boris-specific concepts: Printing Press Software Democratization, Seven Powers Applied to AI. Shared concepts (with Cat Wu / Matt Pocock): Agent Loop Pattern, Harness Shrinkage as Models Improve, Engineer PM Convergence. New entities to support cross-references: Claude Code, Anthropic, Mythos Model. Updated Claude Opus 4.7 (Mythos coexistence + spontaneous-loop behavior at 4.7), Claude Code Best Practices (loop primitive added to Scaling Patterns), LLM-Driven Vulnerability Research (linked to mythos-model entity).

[2026-05-06] ingest | Cat Wu on Lenny's Podcast (How Anthropic's product team moves faster)#

Raw: cat wu how anthropics product team moves faster
Actions: New entity Cat Wu. Cat-specific concepts: AI Native Product Cadence, Claude Character as Product, Model Introspection Feedback. Co-authored shared concepts: Engineer PM Convergence (with Boris), Harness Shrinkage as Models Improve (Cat's to-do-list-crutch case study is canonical). Created Cowork entity; linked Cat as co-product-lead with Boris. Updated Claude Code Auto Mode with Cat's prediction that permission/HITL/static-verification harness assets shrink as models improve.

[2026-05-06] ingest | Matt Pocock workshop (AI Engineer World's Fair 2026)#

Raw: full walkthrough workflow for ai coding matt pocock
Actions: New entity Matt Pocock. Pocock-heavy concepts: Context Window Smart Zone, Design Concept Grilling, Vertical Slice Tracer Bullets, Deep Modules for Agents. Pocock contributes counterpoint to Harness Shrinkage as Models Improve (mechanical verification stays load-bearing) and primary Ralph-loop articulation in Agent Loop Pattern. Updated Agent Harness Engineering (smart-zone constraint, alignment-layer harness primitives, model-introspection debugging, vertical-slice planning invariant) and Claude Code Best Practices (grill-me as aggressive variant of explore→plan→code; deep-modules guidance; cross-links to all four Pocock concepts). Updated LLM-as-Compiler Knowledge Base with Brooks "design concept" parallel.

[2026-05-06] compile | Cross-link sweep#

Actions: Bidirectional backlinks added across the 14 new articles (10 concepts + 4 entity pages) and the 6 updated existing articles (Agent Harness Engineering, Claude Code Best Practices, Claude Code Auto Mode, Claude Opus 4.7, LLM-Driven Vulnerability Research, LLM-as-Compiler Knowledge Base). Three sources cross-reference each other extensively (Boris ↔ Cat ↔ Matt) — captured the convergence around: loop primitive, harness shrinkage, engineer-PM convergence. Index updated: raw 13→16, concepts 12→31, statistics block bumped. Notable orphan candidates (entity stubs) intentional: Anthropic, Claude Code, Mythos Model are central enough to deserve standalone pages even with thin first-pass content; will deepen on future ingests. Deferred again: Agent Context Files cross-vendor sweep (still warranted; not bundled).

[2026-05-06] query | How to learn and co-work with AI as a software engineer (career-skills field guide)#

Output: Learning to Co-Work with AI: A Software Engineer's Field Guide
Actions: Synthesized 14 concept articles into a field guide structured as: mindset shift, six skill clusters (product taste, harness engineering, alignment-first planning, agent-friendly architecture, verification & review, strategic positioning), daily-practices table, anti-pattern table, what-stays-human, 90-day learning plan, source-confidence audit. Heavy reliance on the Boris/Cat/Matt 2026-05-06 ingest cluster plus the 2026-04 harness/best-practices baseline. Filed in derived. Added Derived table entry in Index.md (3→4) and bumped statistics. Added backlinks from Engineer PM Convergence, Printing Press Software Democratization, Harness Shrinkage as Models Improve, AI Native Product Cadence, and Claude Code Best Practices (Derived sections) — these are the highest-density connections for a career-skills-oriented reader entry-pointing into the wiki.

[2026-05-08] ingest | Model Spec Midtraining (arXiv 2605.02087, Li et al., Anthropic Fellows)#

Raw: model spec midtraining improving how alignment training generalizes
Actions: First alignment-research-paper ingest in this wiki. Created core MSM concept Model Spec Midtraining (MSM). Created supporting concepts: Alignment Fine-Tuning (AFT), Agentic Misalignment (AM) (Lynch et al. 2025 eval), Synthetic Document Finetuning (SDF) (Wang et al. 2025 foundation), Deliberative Alignment (Guan et al. 2025 baseline), Model Spec Science (empirical framing of spec authoring), Chain-of-Thought Monitorability (Korbak et al. 2025 fragility argument). Created entity pages Claude's Constitution / Model Spec (Model Spec / Constitution as artifact) and Chloe Li (lead author, Anthropic Fellows). Local PDF saved for appendix lookups. Surfaced cross-domain connection: spec-as-document pattern at Symphony (product spec) generalizes one layer deeper to Claude's Constitution / Model Spec (alignment spec) — both are load-bearing plaintext artifacts, but only the latter is now a direct training input. Backlinks added to Anthropic (Fellows Program, alignment stack), Claude Character as Product (constitution as character substrate), Claude Opus 4.7 (data-generation model for MSM/AFT pipelines), Harness Shrinkage as Models Improve (alignment moves from harness-prompt-injection to model-internalized values), Symphony (spec-as-lever pattern), Claude Code (agentic deployment exposes AM surface).

[2026-05-08] ingest | Why You Shouldn't Treat AI Agents Like Employees (HBR, Kropp et al., May 2026)#

Raw: Research Why You Shouldn’t Treat AI Agents Like Employees
Actions: First workforce-research ingest. Created AI Employee Framing (n=1,261 randomized experiment: AI-employee vs AI-tool framing measurably erodes accountability/escalation/error-catching/role-clarity without lifting adoption), Human-AI Accountability Redesign (five-pillar prescription: span-of-control, role design, performance management, decision-rights/escalation/consequences, agentic-unit-not-human-role), AI Brain Fry (cognitive-fatigue cost surface). Surfaced workforce-side mirrors of existing wiki concepts: Claude Code Auto Mode is a concrete decision-rights instance, Engineer PM Convergence is the engineering-side of "human roles concentrate on supervision/judgment," AI Native Product Cadence is one functional instance of the redesign, Agent Loop Pattern is the output-multiplier that forces accountability redesign, Context Window Smart Zone is the model-side analog of brain fry, Cowork is the deployment surface where these governance choices land outside engineering. Backlinks added in all six.

[2026-05-08] compile | Cross-link sweep#

Actions: Bidirectional backlinks across 12 new articles (9 MSM-batch concepts/entities + 3 HBR concepts) and 11 updated existing articles (Anthropic, Claude Character as Product, Cowork, Engineer PM Convergence, Claude Code Auto Mode, Agent Loop Pattern, AI Native Product Cadence, Symphony, Harness Shrinkage as Models Improve, Claude Opus 4.7, Context Window Smart Zone, Claude Code, Hermes Agent). Index updated: raw 16→18, concepts 31→43. Two new domains opened that previously had thin coverage: alignment research (MSM, AFT, AM, SDF, deliberative alignment, CoT monitorability, Constitution, Model Spec science) and AI workforce/governance (AI employee framing, accountability redesign, brain fry). Both domains intersect existing engineering coverage at deployment surfaces: Cowork, Agentic Misalignment (AM) / Claude Code, and Claude Code Auto Mode are the bridges.

[2026-05-13] ingest | Interaction Models (Thinking Machines Lab, May 2026)#

Raw: Interaction Models A Scalable Approach to Human AI Collaboration
Actions: First Thinking Machines Lab ingest. Created core concept Interaction Models (models that handle real-time multimodal interaction natively rather than via harness). Created supporting concepts: Turn-Based Interface Bottleneck (single-thread turn-taking as a collaboration bandwidth bottleneck; "humans pushed out by the interface, not the work"), Time-Aligned Micro-Turns (200ms interleaved input/output streams; streaming-sessions inference upstreamed to SGLang; latency-tuned MoE kernels; bitwise trainer-sampler alignment), Interaction / Background Model Split (dual-model architecture: present-but-fast interaction model + async deep-reasoning background model; rich-context-package delegation), Encoder-Free Early Fusion (dMel audio embedding, 40×40-patch hMLP, flow head for audio out, all co-trained from scratch), Full-Duplex Interaction (proactive interjection, visual-cue reactions, simultaneous speech, live translation/commentary, time-aware speech), Interactivity Benchmarks (FD-bench + Audio MultiChallenge + new TimeSpeak/CueSpeak and RepCount-A/ProactiveVideoQA/Charades; TML-Interaction-Small 0.40s turn-taking latency, dominates interaction quality). Created The Bitter Lesson as a standalone principle page (Sutton 2019) since it recurs as a load-bearing argument across the wiki. Created entity pages Thinking Machines Lab (the lab) and TML-Interaction-Small (276B MoE / 12B active interaction model). Surfaced cross-domain connection: TML's "harness dissolves into the model" thesis is the same shape as Harness Shrinkage as Models Improve (an Anthropic/Claude-Code observation) — convergent thinking from different labs, both grounded in The Bitter Lesson; and TML's critique of autonomy-first interfaces in Turn-Based Interface Bottleneck is the interface-side mirror of the org-side critique in AI Employee Framing / Human-AI Accountability Redesign. Backlinks added in Harness Shrinkage as Models Improve, Agent Harness Engineering, Context Window Smart Zone, AI Employee Framing, Human-AI Accountability Redesign, Design Concept Grilling, Claude Opus 4.7, Scale-Dependent Prompt Sensitivity, Claude Character as Product, Model Spec Midtraining (MSM), Anthropic. Index updated: raw 18→19, concepts 43→53.

[2026-05-13] query | Opinions on using AI tools & the future of the software engineering role#

Question: What are opinions of using AI tools, and what are the futures of the software engineering role?
Output: Opinions on Using AI Tools & the Future of the Software Engineering Role — debate map of four stances (A bullish-insider: Boris Cherny/Cat Wu — coding-is-solved, harness-shrinks, loops-are-the-future; B pragmatist-practitioner: Matt Pocock — fundamentals apply, bad codebases make bad agents, feedback loops are the ceiling, review is the unsolved problem; C skeptic-governance: HBR/Kropp et al. AI Employee Framing/AI Brain Fry — anthropomorphizing erodes accountability −9pp/escalation +44%/errors caught −18%, no adoption gain; D architecture-thesis: Interaction Models/The Bitter Lesson — harness dissolves into the model, interaction included). Part 2 synthesizes the future SWE role: coding→deciding-what-to-build + designing-the-agent-environment + verifying; role convergence & smaller teams; what stays human (code review, EQ, domain depth, mission clarity, system-level oversight, verification infra); which moats survive (process power & switching costs erode; network effects/scale/cornered resources persist; counter-positioning amplifies); honest caveats (most evidence is Anthropic-internal + one practitioner; "100 lines" is self-described hyperbole; the one randomized study is a warning not an endorsement).
Actions: Filed derived page. Added Derived backlinks in Printing Press Software Democratization, Harness Shrinkage as Models Improve, Engineer PM Convergence, Seven Powers Applied to AI, AI Employee Framing. Cross-linked as the descriptive companion to the prescriptive Learning to Co-Work with AI: A Software Engineer's Field Guide. Index updated: derived 4→5, last-derived 2026-05-06→2026-05-13.