Engels' Pause?

What we talked about

Knowledge graphs

Came up as the structured counterpart to raw retrieval — entities and the typed relations between them, so a model can traverse “who connects to what” instead of guessing from adjacent text. The pull is grounding: an explicit graph gives an LLM a fact substrate it can walk, and a place to catch contradictions. Open question for the room — where does building the graph actually pay off versus just throwing more context at the model? — Knowledge graph — Wikipedia — Neo4j — What is a knowledge graph? — Microsoft GraphRAG

Graph databases

The storage layer under the above — Neo4j, FalkorDB and friends, where nodes and edges are first-class and a query is a traversal rather than a pile of joins. The interesting bit is the pairing with agents: an LLM that can emit Cypher (or read a graph summary) turns “explore the neighbourhood of this entity” into a cheap, deterministic lookup. When does a graph DB earn its keep over a plain relational schema? — Neo4j — FalkorDB — Graph database — Wikipedia

Vector databases

The other half of the retrieval conversation — embeddings and nearest-neighbour search for “find me things that mean roughly this.” Contrasted with graph DBs as fuzzy-vs-precise: vectors are great at semantic recall and terrible at exact multi-hop relationships, graphs the reverse. The live debate was hybrid retrieval — vector search to find the entry points, graph traversal to walk the structure from there. — pgvector — Pinecone — Weaviate — Qdrant

Perseus

Lettria’s pipeline for turning documents (or raw text) into a knowledge graph — extract entities and relations, optionally constrained by an ontology, then push the result into Neo4j or FalkorDB. Sits right at the intersection of the three topics above: it’s the “documents → graph” on-ramp, with an MCP server so an agent can drive the whole build-and-export flow. — perseus.lettria.com

ElevenLabs and AI voice technologies

The voice side of the table — ElevenLabs for high-fidelity TTS, voice cloning, and increasingly real-time conversational agents. Discussion ranged from the obvious uses (narration, dubbing, accessibility) to the agentic ones: giving a coding or research agent a voice, or wiring speech in/out so you can talk to a workflow instead of typing. Open question — where does voice genuinely beat text as the interface, and where is it just a novelty? — elevenlabs.io

Claude vs Claude Code vs Codex

Untangling three things that get lumped together — Claude the model, Claude Code the agentic harness wrapped around it, and Codex as OpenAI’s competing harness. The point that landed: the harness matters as much as the model, since it owns the tools, the context, and the loop. Round-the-table on who’s running which, and whether people pick by model quality, by harness ergonomics, or just by what their employer pays for. — Claude — Claude Code — OpenAI Codex

Getting AI to write better design docs

How to make an agent produce a design doc worth reading instead of a padded restatement of the prompt. Tactics that came up — make it interrogate you first (grill the assumptions before drafting), ground it in the actual codebase rather than generic best-practice, force explicit trade-offs and rejected alternatives, and keep it at the right altitude instead of drowning in implementation detail. Open question — what’s the prompt or workflow that reliably gets you a doc you’d actually hand to a teammate? — Design Docs at Google — Anthropic — Prompt engineering

llama-swap (demo)

Someone demo’d llama-swap — a lightweight proxy that sits in front of llama.cpp (and other OpenAI-compatible backends) and hot-swaps the loaded model on demand, so one endpoint can serve many local models without you babysitting which is loaded. Came up as the practical glue for running a stable of local models on limited VRAM — request a model, it loads it, unloads the last one. Handy for keeping a local coding/embedding/voice stack on one box. — github.com/mostlygeek/llama-swap

Claude Code dynamic workflows

Revisited the dynamic workflows Anthropic shipped recently — Claude writes a JavaScript orchestrator from your prompt and the runtime fans it out across up to 1,000 background subagents, with the script (not the chat context) holding the intermediate state. Tied back into the day’s threads: it’s the obvious engine for “documents → knowledge graph” builds and other fan-out jobs. Note the trigger keyword moved from workflow to ultracode in v2.1.160. Where does it actually beat /goal + /loop, and what’s the real token cost? — Anthropic — A harness for every task

Engels’ pause

Brought up as the historical rhyme for the moment we’re in — the stretch of early-19th-century Britain (roughly 1790s–1840s) when industrial output and productivity climbed steeply but real wages for workers flatlined for decades before finally catching up. The parallel on the table: AI may be opening a similar gap now, where the gains land with capital and the labour side waits a generation for the lift. Open question — is this time a pause that resolves the same way, or a structural break, and what shortens the lag? — Engels’ pause — Wikipedia

Remotion

The “video as React” angle — write components, drive them with props and a frame counter, and render the timeline to MP4 with a headless browser. The agentic pull is that video becomes code an LLM can author and diff: parametrise a template, feed it data, batch-render variants. Where does programmatic video actually beat a timeline editor — data-driven explainers and per-user clips, or is the authoring overhead only worth it at scale? — remotion.dev — github.com/remotion-dev/remotion

HeyGen

The avatar end of AI video — script in, a talking presenter (cloned or stock) out, with translation and lip-sync across languages. Came up alongside ElevenLabs as the visual counterpart to AI voice: between them you can stand up a synthetic presenter from text alone. Open question — is this genuinely useful for docs, onboarding, and localisation, or does the uncanny-valley tax still outweigh the time saved? — heygen.com

hyperframes

HeyGen’s open-source take on controllable AI video — a framework for stitching generated frames into coherent sequences rather than one-shot clips. Interesting as the hackable layer under the polished product: if the frame-level control is real, it’s the bit you’d wire into an agentic pipeline. Worth a look at how much steering it actually gives versus the hosted product. — github.com/heygen-com/hyperframes

Timestamp-based cutting

The editing trick that ties the video thread together — transcribe a recording, then cut by editing the transcript: delete a sentence, the corresponding clip drops out. Pairs naturally with the AI-voice and avatar tools above, and it’s squarely agent-shaped since the model works on timestamped text, not pixels. Where’s the line where this replaces a real editor versus just roughing out a first pass?

Cline

Back to coding harnesses — Cline as the open-source, VS Code-native agent, often paired with whatever model you bring (including local backends via something like llama-swap). Slotted into the Claude Code vs Codex comparison as the BYO-model, fully-in-editor option. Round-the-table on who’s running it, and whether open + editor-native wins out over the polish of the hosted harnesses. — github.com/cline/cline — cline.bot

Announcements — 11:55

• OKTech — June event — Date, time & venue TBD — details and signup at oktech.jp.
• RSVP reminder — Coming along? RSVP on Meetup so we can size the room — meetup.com/oktech.
• Feedback — Got thoughts, topic ideas, or want to demo something? Drop us a line — feedback.

News since last assembly

Floor: 2026-05-30 (aa03). Generated 2026-06-06.

New Claude Code commands & features

• ultracode trigger rename (v2.1.160, 2026-06-02) — the dynamic-workflow keyword we covered last week is no longer workflow; type ultracode in a prompt (or /effort ultracode) to auto-orchestrate a workflow — release
• /plugin list (v2.1.163, 2026-06-04) — list installed plugins with --enabled / --disabled filters — release
• claude agents fan-out progress (v2.1.161, 2026-06-02) — shows done/total for fanned-out work, and --json now reports a waitingFor field (v2.1.162) — release
• Safer writes (v2.1.160, 2026-06-02) — prompts before writing to shell startup files / git config; acceptEdits now confirms before touching build-tool config — release
• Autocomplete no longer auto-runs (v2.1.162, 2026-06-03) — picking a slash command from autocomplete fills the input instead of firing immediately; Remote Control is now a persistent footer pill — release
• SubagentStop additionalContext (v2.1.163, 2026-06-04) — SubagentStop hooks can return hookSpecificOutput.additionalContext to feed text back to the parent — release

Codex

• /archive + codex archive (0.136.0, 2026-06-01) — archive/unarchive sessions from the TUI or CLI; TUI markdown keeps clickable links via OSC 8 — release
• [2026-06-04] Multi-agent v2 preserves runtime choice per thread; web/image tools gain parallel search in code-mode; enterprise builds show monthly credit limits (0.137.0) — release
• [2026-06-01] Windows alpha codex sandbox setup --elevated provisioning path; remote execution via CODEX_API_KEY for approved hosts — release

Adjacent tools

• [2026-06] Cursor adds canvas Design Mode + interactive context-usage report, and a Premium seat (5× usage) in a Teams pricing rework — changelog
• [2026-06-03] VS Code 1.123 ships more agent features, larger model-context support, and integrated-browser updates — release notes
• [2026-06-02] Microsoft unveils MAI-Code-1-Flash at Build 2026 — its first in-house coding model, not OpenAI-powered — Simon Willison

Simon says

• [2026-06-03] Uber caps employee AI-tool spend at $1,500/mo after burning its 2026 AI budget in four months — post
• [2026-06-04] “AI enthusiasts are in a race against time, AI skeptics are in a race against entropy” — post

Notable posts

• [2026-06-01] GitHub Copilot switches every plan to usage-based AI Credits; power users post projected 10–50× cost jumps; the community thread tops 400 comments / ~900 downvotes — discussion

Topics worth a 5-min slot

1. The metering era — Copilot AI Credits, Uber’s $1,500/mo cap, and Cursor’s reworked Teams pools all landed the same week. Has agentic coding’s “all-you-can-eat” phase ended?
2. workflow → ultracode — last week’s dynamic-workflows trigger got renamed in v2.1.160. Worth a live demo now that the keyword is stable.
3. In-house coding models — Microsoft’s MAI-Code-1-Flash drops the OpenAI dependency. Does owning the model matter when the harness is the product?