From scrubs to swarms: the two-year arc of agentic work at 8am AI

Most of what gets said about AI is a snapshot — a benchmark, a launch, a take. The 8am AI corpus is unusual because it's a time series: the same small group, most Wednesdays, for two years. And the single thread that runs cleanest through it is the one about agents — how you build them, what you point them at, and who's actually in charge.

Here's the arc, grounded in the meetings it came from.

2024: agents you build by hand

The earliest sessions (April–May 2024) open with a word the group used before the industry settled on "agent": scrubs. A scrub was a basic set of instructions — a small, functional unit you'd give an LLM to make it do one job reliably. The emphasis from day one was self-reflection: getting the model to check its own output to improve responses.

By summer 2024 the demos get concrete. David builds a product-description agent and runs it against a real project (Hanif's "Stonemaster"). The group watches agents collaborate, watches one agent's output become another's input, and starts talking about simulation — agents standing in for people to see how a workflow behaves.

The defining constraint of this era: everything is hand-wired. One agent, one prompt, one task. The work is in the authoring.

2025: orchestration replaces authoring

Through 2025 the conversation moves up a level. The recurring topics stop being "how do I make an agent" and become "how do I wire agents together":

• N8N for connecting APIs and automating multi-step flows.
• Local model runners and MCP servers, so agents can reach tools and data.
• Workflow builders — node editors where a flow is assembled visually rather than written out prompt by prompt.

This is where WorkSona shows up in the corpus: a lightweight, portable MCP server with standardized agents and reporting, built so a small team can stand up an agent by dropping a JSON file in a folder. The framing shifts from I wrote an agent to I orchestrated a workflow — and, tellingly, to I can hand this to a non-programmer teammate.

The honest counter-current runs alongside it: AI-generated code violating an AutoCAD license, context windows corrupting mid-task, the now-famous "you're absolutely right" as a tell that the model has lost the thread. The group never treated the tools as magic; the failure modes are logged as carefully as the wins.

2026: the agentic layer, and the swarm

By January 2026 the ambition is stated flatly: adopt an agentic layer in all work this year. Not "use agents for a task" — make the derivative, after-hours, agents-finish-the-job layer a default property of how work gets done.

Then, in March 2026, the inflection the group had been circling for two years arrives in the conversation: OpenClaw and autonomous swarms. The discussion is striking for how matter-of-fact it is. Developers describe taking a director role — commanding fleets of agents, delegating observation and analysis, managing risk and validation rather than doing the atomic work. The same critical-thinking discipline persists, the group argues, just at a higher altitude. The worry — that there'll be "no ladder to climb" for people who never did the hands-on work — is named and debated, not waved away.

Two years on, the unit of work has inverted. In 2024 a human authored an agent to do a task. In 2026 a human directs a swarm and the agents author the work — sometimes, in the group's own projects, writing their own code and their own tickets.

What the arc teaches

Three things only the time series makes visible:

1. They were early. David was running collaborating agents against real projects in mid-2024 — roughly a year before "agentic" became a headline. The corpus is a receipt for that.
2. The hard part moved. In 2024 the difficulty was authoring the agent. In 2025 it was orchestration. In 2026 it's judgment — knowing what to point the swarm at and how to validate what comes back. The work didn't disappear; it climbed.
3. The discipline held. Across every era the group's through-line is the same: self-reflection in 2024, observability in 2025, validation in 2026. Different words, same insistence — the human stays in the loop as the evaluator, no matter how much the machine takes over.

Open question

If the agents are now authoring the work, the scarce human skill is deciding what is worth building and whether the result is true. The corpus has a name for the second half of that — "how do you know" — and not yet a good answer. That's the thread to follow next.