Measure
Write the success check, stop condition and reproduction path first.
AGI SYSTEMS / PUBLIC LAB
Before claiming AGI, inspect the loop.
Observe the goal, make a plan, take the smallest reversible action, verify the result, then change the next move. Intelligence is more credible as a loop that another person can follow than as a grand claim.
This demo runs in your browser. It does not send input to an external model or API. Do not enter confidential information.
LOOPinspectable
01OBSERVEthe state 02
PLANthe move 03
ACTsmallest change 04
VERIFYthe evidence 05
LEARNnext loop
Ship
Release the smallest working artifact that another person can review.
Fix
Change one variable, record what happened, and return to the loop.
LOCAL LOOP ENGINE
Run the five stages with your own mission
Your input stays in this tab. This is a transparent sandbox for control-loop design, not a claim that the page itself is AGI.
LIVE TRACE
An execution log you can inspect
Ready
01OBSERVEstateidle
02PLANmoveidle
03ACTartifactidle
04VERIFYevidenceidle
05LEARNnextidle
ITERATION ARTIFACT
RUN 01The next move
PLAN
STOP IF
NEXT LOOP
SYSTEM MAP
Design the boundaries, not just the model
A real agent needs more than a capable model. Separate tools, evaluation, memory and permissions so failure can be stopped and understood.
Loop controller
Accept a mission, manage the next stage and enforce the exit condition. Never run forever by default.
state → action → evidenceReasoning
Generate hypotheses. Treat the output as a hypothesis, never as evidence.
Tools
Search, code and calculate only inside an explicit permission boundary.
Evaluator
Compare the result with the success check and record the failure mode.
Memory
Decide what to keep, when to forget it and who is allowed to read it.
Permissions
Stop irreversible actions, personal data and cost-bearing actions until approved.
WHAT WE MEASURE NEXT
Do not optimize for looking intelligent
Public work needs comparable measures, not vibes. These are measurement plans, not claimed results.
Useful artifact latency
Time from the mission to an artifact another person can review.
Verifier pass rate
The share of iterations that satisfy the success check written in advance.
Rollback rate
The share of failures that can be safely undone. Being able to stop is part of capability.
Operator load
How many checks, corrections and approvals the human had to provide.
PUBLIC EVIDENCE
Move from claims to working evidence
Use the existing small implementations to test the ideas behind this lab.
OPEN LOOP
Bring the next test.
Technical feedback, a review of the public demo or a prototype discussion is welcome. Do not wait for a scout; keep shipping work that is worth finding.
Send technical feedbackWhat this page is
AGI Loop Lab is a public workbench for seeing what an "AI agent" actually does, broken into five stages: observe → plan → act → verify → learn. Enter a goal and constraints in the loop engine above and trace how the five stages unfold. It is not connected to any external AI service; the output is a mock generated by rules inside this page. The point is to understand the shape of the loop before building a real agent.
How it works
- Loop control: take a goal, run the five stages in order, stop when a STOP IF condition is hit. It never runs forever.
- Evidence: the output of "plan" or "reason" is not treated as evidence; only what passes the "verify" stage against the success criteria feeds the next loop.
- Permission boundaries: irreversible actions, personal data and anything that costs money are where a real agent must stop for a human. The demo marks those spots.
How to use it
- Write the goal in one sentence (e.g. automate the weekly report).
- Write the constraints: what must not happen (e.g. never modify production data).
- Press "Run the loop" and read the five stages in LIVE TRACE, plus "next move" and "STOP IF."
- Change one constraint and run again. One change at a time is the house rule.
Where your data lives
The goal and constraints you type are used only inside this tab; nothing goes to a server or an external AI, and it is gone when you close the page. Even so, do not enter personal, secret or unreleased business information.
FAQ
Is this a real AI?
No. There is no language model behind it. It is a prototype for experiencing the shape of the loop; the output comes from fixed rules.
What is it good for?
When you build your own agent, it trains the habit of deciding success criteria, stop conditions and permission boundaries before worrying about how clever the model is. The SYSTEM MAP above is the blueprint for that design.
Where can I see a real agent?
Every work on this site was built with a coding agent. How is described in How this site was built with Claude Code; day-to-day field notes are at AGENT NOTES.