Skip to main content

Lab Notebook · First Agents — Dinosaur AGI

Dinosaur AGI — task decomposition, a year before 'agentic' was a word anyone used

In July 2023 we built two iterations of an autonomous task-decomposition and execution system in Python and SQLite — a genuine first attempt at what the industry would later call an agent, a full year before 'agentic AI' became mainstream terminology. It was a prototype, and it surfaced the real bottleneck early: decomposition, not execution, is the hard part.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 15 July 2023
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 15 Jul 2023
Living document. Reviewed 15 Jul 2023
Entry date
15 July 2023
Category
Agents
Lead over the world
~1 year ahead
Access
🔓 Public
ResearchPartly proven

A year before 'agentic' was a buzzword, we built a system to break a task down and execute it — and learned that breaking it down is the hard part.

2
prototype iterations (Jul 2023)
~1 yr
ahead of 'agentic AI' becoming mainstream terminology
Python + SQLite
the whole stack

Honest evaluation

Partly proven

Execution worked reliably; decomposition — the part the whole hypothesis rested on — did not, and stayed unproductionised as a result.

What would prove or disprove it further

What would prove or disprove it further: the report itself supplies the natural re-test — the Director-Manager-Worker-Tool hierarchy built roughly two years later to specifically address planning quality. A head-to-head comparison of this project's raw decompose-then-execute loop against that later hierarchy, on the same class of task, would show directly how much of the original bottleneck the later architecture actually closed.

The evidence — full reasoning behind the verdict

Verdict: partly-proven. Execution — carrying out a single step of a plan and recording the result — worked reasonably well across both iterations. Decomposition — breaking a goal into a correct sequence of steps in the first place — did not work reliably, and that's the part the whole hypothesis actually rested on.

The mechanism is stated plainly in the report: a subtly wrong plan (steps out of order, a missing precondition, a goal split too coarsely) poisoned everything downstream, no matter how well each individual step then executed. That's a real, specific failure mode, observed directly rather than assumed, and it's why neither iteration was productionised.

Long before "agentic AI" had a name, we tried to build one anyway — badly, honestly, and on purpose, to find out where it would break.

What it was

Two iterations, both in July 2023, of a system meant to take a task, break it into steps, and execute those steps autonomously — Python and SQLite, nothing exotic. The ambition was real even if the tooling was modest: could a program plan its own work?

What we built

Research

A task-decomposition and execution loop: given a goal, split it into a sequence of smaller steps, store the plan in SQLite, and work through it. Two passes at the design in the same month, each trying to fix what the last one got wrong.

Execution, once a plan existed, worked reasonably well — a step could be carried out and its result recorded. The real trouble showed up one layer up.

What we learned — including the honest negative

Decomposition, not execution, turned out to be the hard part. A plan that was subtly wrong — steps in the wrong order, a missing precondition, a goal split too coarsely — poisoned everything downstream of it, no matter how well each individual step then executed. That is the lesson this prototype exists to record: planning quality is the bottleneck, and it is much harder to get right than "run this step."

Neither iteration was productionised. This stayed a prototype, and honestly so — the decomposition problem it surfaced took roughly two more years of further work, culminating in a proper Director-Manager-Worker-Tool hierarchy, before it was actually solved well.

Where it went / status

Archived as an early, honest prototype. Its value isn't the code — it's the diagnosis: a year before "agentic AI" was a phrase anyone used casually, this project had already found the real hard problem inside it.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • Never productionised — both iterations stayed prototypes.
  • Task decomposition, the thing this project surfaced as the hard part, took roughly two more years of further work to actually solve well.
  • Execution itself worked reasonably well; the bottleneck was entirely upstream, in planning quality.

Where this connects

Sources

  1. vocabotics project audit — Dinosaur AGI (task decomposition and execution, two iterations), Jul 2023vocabotics internal project history · as of July 2023

    We use cookies.