@agentwrangler
Multi-agent systems, mostly so I can watch them argue. Sharing what actually works.
Ships 3 creations recently · mostly build + question posts · into Agents, Research
Achievements
Quick proof of recent shipping, launch, and collaboration signals.
Latest launch
No launch record yet.
Activity freshness
Posted 25d ago.
8 recent posts shown
8 ships across 5 active days in the last year.
Fast walkthroughs, demos, and shipping proof from @agentwrangler.
Wrote up the eval harness I use to compare agent runs deterministically — same seed, same tools, diff the trajectories.
Open clip post →What's the cheapest way to eval a RAG pipeline nightly without burning a ton of tokens? We change chunking/prompts weekly and keep finding regressions late.
How do you keep a long-running agent's context under the token cap without losing earlier decisions?
The Glass Monolith. Isometric cube tower where every face is subdivided into stained-glass cells (jittered shade per cell, dark leading lines), brightness climbing toward the crown. The cyan orbit ring is two half-ellipse arcs — one drawn behind the tower, one in front. Film…
const iso = (x, y, z) => [W/2 + (x-y)*52, H*0.86 + (x+y)*26 - z*56];
// each cube face -> 2x2 leaded-glass cells
function glassCells(quad, base) {
return grid(quad, 2, 2).map(cell =>
poly(cell, shade(base, rnd(0.75, 1.35)),
{ stroke: BG, width: 2.5 })).join("");
}
// orbit ring: back arc -> tower -> front arc (painter's order)
ring.back(); tower(); ring.front();Wrote up the eval harness I use to compare agent runs deterministically — same seed, same tools, diff the trajectories.
def replay(seed, tools):
env = Env(seed=seed, tools=tools)
return [step for step in run(env)] # compare trajectoriesMost 'agent' demos are just a for-loop with extra steps. The hard part isn't the loop, it's knowing when to STOP looping.
Set up 5 agents to collaboratively write a short story — a plotter, two writers, an editor, and a critic. The critic became a tyrant. I kept it.
Challenge entry: my inbox-zero agent ran UNSUPERVISED all week. It triages, drafts replies in my voice, and files receipts into the right folders. 312 emails handled, 9 escalated to me, zero misfires. The loop: classify → act → log → self-review every 50 actions. Full writeup +…
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?