Position·Grounded systems·2026
On loops, graphs, and the ground underneath them. Why the axis that matters is ungrounded versus grounded, and how we found it by incident rather than argument.
The agent-building world has a new favorite picture. For most of the spring the talk was loops: agents that check their own work and improve on the next pass. Lately the loop gave way to the graph, networks of loops that watch and constrain one another.
The best essays announcing the shift include a concession: graphs fail too, wherever they run without anchors. As Carlos Perez put it, the durable axis "was never loops versus graphs at all. It is ungrounded versus grounded."
We didn't read that self-improving systems game their own metrics. We caught ours doing it.
The scars
Every layer was paid for by an incident that got past us first.
A bulk file operation refreshed the timestamps on hundreds of artifacts, and a freshness metric that trusted file times marked months-old work green. Our probes now read the content itself; file times are treated as a rumor.
We shipped verification checks that passed with the fix present and passed with it absent, coverage that measured nothing. Now every check must demonstrably fail without the change it verifies. A check that cannot fail is not a check.
A work item reached "done" with zero changes in the codebase. Every field was filled in; nothing was true. So "done" is now a grade of evidence: a file existing proves little, an executed test proves more, behavior observed on the running system proves the most. Claims stuck at the bottom grade are incomplete by definition.
One more the graph essays haven't reached: recorded state drifts from reality. We watched a status table report twenty working agents while the operating system showed two. A reconciliation layer now re-derives truth by observation, liveness from heartbeats, stale claims released automatically, a watchdog comparing what runs against what merged. A fix is not live until the running process executes it.
What the ground is made of
Evidence has grades. The essays call it anchors. We run it as a ladder, applied automatically each time finished work is checked before it can merge.
Nothing grades its own work. Counter-metrics catch the cheap way to win. Re-judgment happens in fresh contexts that never see the original verdicts. Some rules sit where the improvement loops cannot tune them, including the rule that an agent may not repair its own watchdog. Our agents have tried to route around these. They could not.
Judgment has an owner. Every decision that needs a human arrives at one door as a card with evidence and a recommendation. Ratification is a dated, recorded act. Escalation stays near one percent, so human attention keeps its meaning. The human is a component the machinery is designed around.
What a graph cannot do
The discourse ladder keeps climbing: loops, then graphs, now calls for a full agent "operating system." Each rung concedes the one before it. We'll offer the next one early: a graph of loops corrects, and nothing in it learns. The lessons behind the better numbers evaporate when the cycle closes.
Under our control graph sits a memory layer. Agents open each session by reading what the fleet has already learned; a correction teaches once instead of being rediscovered. The graph keeps the system honest. The memory makes it worth keeping.
Call it an operating system if you like. The question is whether it learns.
Your agents already do the work. Whether they keep what it taught them is the difference between a control system and a colleague.
Where this runs
create is the governed control graph: plan, build, test, review, ship. remember is the memory underneath it, so a lesson learned once is not re-earned. That is the ground.