Two years of 8am AI: what a weekly conversation actually recorded

Since April 2024, a small group has met most Wednesday mornings at 8am to talk about what's actually changing in their work as the models improve. Two years and 109 recorded sessions later, the substrate has settled into fourteen recurring themes and roughly 760 distinct ideas — mined not from a plan but from the transcripts themselves.

This is the first look at the whole corpus end to end. It started as "8am AI School" — workshops with numbered sessions — and became, around mid-2025, just "8am AI!": a standing peer conversation. The name change tracks a real shift in what the meetings were for.

The single clearest arc: agents

If you follow one thread across the full two years, follow the agentic one. It's the largest cluster in the corpus (143 ideas) and the only one with a clean, legible trajectory:

• 2024 Q2 — "scrubs." The earliest sessions are David introducing scrubs: basic sets of instructions that turn an LLM into a functional little agent. Agents are built by hand, one prompt at a time.
• 2024 Q3 — agents in spreadsheets. Product-description agents, agent collaboration demos, agents reading and writing Google Sheets. The unit of work is a single task automated.
• 2025 Q1–Q2 — orchestration and tooling. N8N, MCP servers, workflow builders. The conversation moves from "make an agent" to "wire agents together." WorkSona appears as a local MCP server with standardized agents and reporting.
• 2026 — the agentic layer everywhere. By January 2026 the framing is "adopt an agentic layer in all work this year." By March, OpenClaw and autonomous swarms; people describe taking a director role over fleets of agents rather than doing the atomic work themselves.

Two years compressed: from writing one prompt to directing a swarm. The group lived the whole curve in public, week by week.

The four threads that carried the most weight

By volume of ideas across the corpus:

1. Agentic workflow (143) — above.
2. Context engineering (97) — the recurring discovery that the model is the commodity and the context is the lever. Starts in 2024 with "scrubs" as instruction sets; by 2026 it's full corpus-building, the project-state pattern, and "how do you know" as the central question.
3. Product & business (61) — the slow turn from "AI as a product" to "AI as a process," value-based pricing replacing SaaS seats, and the repeated claim that access to the market — not capability — is the scarce thing.
4. Developer productivity (59) — a two-year log of tool churn: GPT-4 → Claude-in-Canada → N8N → Cursor → slop-code culture → multi-agent orchestration. Also the honest counter-current: AI-generated code violating licenses, context corruption, "you're absolutely right" as a failure signal.

Close behind: system-vs-design thinking (59), the move from atomic execution to architecture; and meta/meeting-cadence (54), the group repeatedly re-examining how it should meet as the work changes.

The threads that stayed small but sharp

• LLM evaluation & trust (14) — "they don't know how they know," your input is the eval, Git-gists-as-data-substrate for verifiable records.
• Hiring & authentication (12) — Blade Runner 2049 as a lens on interviewing candidates who run an LLM co-pilot during the interview.
• Future & society (15) — Neuralink, Ozempic-as-norm, UBI, the erosion of community, AI companions. The corpus's philosophical edge.
• Project-state substrate (8) — small in count, large in influence: the structured-memory pattern that this very facility is built on.

What two years of data is good for

The point of keeping the corpus isn't nostalgia. Three things fall out of it that no single meeting could give you:

1. Trajectory. You can see which ideas were early (David doing agents in spreadsheets in mid-2024, a year before "agentic" was a headline) and which were reactions to the news cycle.
2. Recurrence. Themes that keep returning despite no one deciding to raise them — context engineering, the atomic-to-architecture shift — are the ones worth building on. The substrate surfaces them mechanically.
3. Provenance. Every idea cites the meeting it came from. The claims in this post are traceable back to specific Wednesdays.

What's next

This is the index. From here the facility works the queue — 228 graded ideas, top-of-list first — turning the highest-signal threads into deeper pieces. The first deep-dive, on the two-year arc of agentic work, is the companion to this post. The book isn't written; it's being assembled out of what was already said, one Wednesday at a time.