All thoughts

Learning the Agent OS (9 Months)

I stopped chasing model launches and started drawing the system: memory, harness, loop, eval, then System 1 next to System 2. Three short parts, two stages each, with a small drawing for every stage.

How it works

Nine months, three partsFour blocks joined by arrows, one per part and one for now. Only the last block is gold.1. Foundations: Day 1 to ~175, launches and memory. 2. The harness: Day ~185 to ~190, loop and eval. 3. Own system: Day ~265 to ~270, your harness, System 1 and 2. 4. Practicing: Day ~285, the spine I use now (highlighted). 01FoundationsDay 1 to ~175,launches andmemory02The harnessDay ~185 to~190, loop andeval03Own systemDay ~265 to~270, yourharness, System1 and 204PracticingDay ~285, thespine I use now
Select a step to see what it does.
The arc from late 2025 to now. Each part adds two stages; the gold block is where it lands today.

This is nine months as a series, from late 2025 to what I'm shipping now. Each stage added one piece to the system, and each part covers two stages. The drawings grow with it, and only the piece that stage switched on glows. Here's where it ends up, then the parts in order.

Day ~285 (now): the full stackSix platforms from bottom to top: memory and skills, agents with tools and a loop, trace, the eval gate with its System 1 check, a human gate, and Ship. The trace screen, the eval gate and the Ship deck are lit in amber, the same as the factory drawing in the pstack post.diagnose → fixMemory + skills3 pillars + gateAgentstools + loopTracewhat ranEval gatePass to shipHuman GOapproveShiprelease
Where it ends up, read bottom to top. Each stage in the parts below adds one of these pieces.

The series

  1. Foundations: Stop chasing launches (Day 1, Dec 2025) and Give it a memory (Day ~175, Jun 2026).
    Apps beat model theater. Memory is the moat.
  2. The harness: Put a harness on it, and a stop on the loop (Day ~185, Jun 2026) and Trace it, eval it, then ship (Day ~190, Jul 2026).
    A harness rides the LLM horse, and every loop needs an end condition. "Self-improving" without eval is cosplay.
  3. Your own system: Draw your own harness (Day ~265, Sep 2026) and Put System 1 next to System 2 (Day ~270, Sep 2026).
    Cool demos aren't use cases, and portable memory beats locked silos. Fast judgment enhances slow reasoning. It doesn't replace it.

The whole thing on one page

Learning the agent OS: worldview, memory, harness, loop, LLM Ops, personal harness, System 1+2, with a row of checks underneath

The same stages as one spine. Teal is practice I keep. Coral dashed is a runaway loop with no eval. Hold ship until Pass.

My take

I used to refresh model launch posts. Then I spent nine months drawing boxes instead.

Day ~285 · Oct 2026 What I'm practicing now

Stacked up, it's the same factory I draw in How I Use pstack: memory at the bottom, agents in the middle, checks and a person before anything ships. Day to day that means:

  • Clear diagrams on this site
  • Grok Bot as an agent OS
  • Memory and loop engineering written up here, not as slogans
  • Manual first, then routines
  • Source of truth over remembered numbers

What I learnedNo guarantees. Just the spine.

Try this

  1. Pick the part that matches where you are and do its Try this first.
  2. Draw the system you use today as boxes once: memory, harness, loop, eval. Mark the one piece you'd switch on next.

Sources