Run settings — the loop's safety limits (yours to set)
loading…
The loop right now
The Agentic Loop in motion
how a work item flows · who writes/reads what · both CSVs maintained by the loopwatcher tooling · read-only · LD-13
Design visual (the sim-watching surface — sims feed into this HTML, R6f). Corrected Rev-3 design: 8 blocks · THEME-artefact lifecycle · kind-conditional outputs · derived board. Two views (pipeline · radial HUD) show the SAME data. Every theme runs the SAME pipeline; only the output shape varies. Both loops iterate: Planner⇄Plan-critique (SOUND) and Generator⇄Evaluator (the climb).
⚠ SAMPLE DATA — the animation below plays illustrative design scenarios (sims you can play any time). The specific sample scores (quality (F3) 92, quarantine 38→90) are design values. REAL runs — read from actual run data on disk — appear in the green panel just below; pick one to see its real numbers.
✓ REAL RUNS (from disk) — the sims above stay playable; these are actual runs with real quality scores & interference counts.
OVERSEERread-only watch over the whole loopno live run — the Overseer is watching for the next one
REVISE ⇄ re-review
quarantine ⇄ re-build (the climb)
Done? grading
OUTPUT → product board (per output_kind)
Radial HUD PLAYS the scenario: the active block wedge glows and the playhead sweeps to it · inner ticks = the 33 loop-board columns filling · outer ticks = the 44 product-board columns · artefact labels glow amber when WRITTEN / violet when READ · the core ring turns green PASS / red quarantine · click a wedge or artefact to unfold + draw its producer→reader edges.
×
Ring
Per-stream swimlane — one row per stream of the selected real run · block status in plain language
Build kanban — what this run planned / is building / has graded
Planned 0
In progress 0
Done / graded 0
Simulation suite — the loop's 18 self-tests (loop-authored, the simulation phase (PH2)) · this HTML is the sim-watching surface
Each sim is a real cycle with a disk-checkable PASS. A lit chip = a scenario above DEMONSTRATES that sim's mechanic; click it to play that scenario. A dim chip = authored + queued, not yet animated here. Hover any chip for its PASS condition.
How the loop improves ITSELF FIRST-CLASS · every cycle
Not a footnote — evolution is a stage of every run. Dreams route into the next plan, agent/skill proposals queue (Channel B), reflections peer-review, the north-star GOAL sharpens, and the bar only ever rises (monotonic). These light up as the run reaches them.
Every artefact — who PRODUCES it, who READS it — explicit producer→reader edges (R6f.1/.3)
The complete set the loop touches (loop-native + machinery + product family incl. US / user-story). Click a row to spotlight its producer + readers in the view above and draw the edges on the HUD.
Artefact
Produced by
Read by
Template / where
Both CSVs — which block writes which column — explicit block→column map (R6f.2)
Every cell is tagged + colour-coded by its writing block (left border). Click a block below to spotlight exactly the columns it writes across BOTH boards. During Play the active block flashes the cells it writes.
agentic-loop-runs.csv — loop board (33 cols) DERIVED from per-stream state.csv
quality trend (F5) needs ≥2 runs — pending 2nd run
📈 Model rate limits & health — MEASURED from real fires
One colour-coded line per model the loop fires — its solid line is successful fires per time bucket; amber-ringed dots mark rate-limited (429) buckets and red-ringed dots mark provider brownouts (503). Below the chart, each model row shows its numbers and which blocks are pinned to it, so a switch decision is one glance. A model with no events says so — nothing here is invented.
loading…
how to read the chart:
a model's solid line = its successful fires per bucket
rate-limited (429) at that bucket
brownout (503) at that bucket
— dots take the model's fill colour, so you can tell whose trouble it was