LoopGain @LoopGainAI

Open-source cost control for AI agent loops — stop paying to run loops past the point they've already converged. Stops at convergence, rolls back on divergence. loopgain.ai Toronto, Canada Joined May 2026

Tweets

14
Followers

6
Following

125
Likes

7

LoopGain @LoopGainAI

14 hours ago

loopgain.ai/blog/posts/how…

0 0 0 1 0

View Details

LoopGain @LoopGainAI

14 hours ago

We re-graded our own benchmark against a fuller test suite and found ~1 in 20 "converged" runs were actually wrong — a bug on an edge case that wasn't sampled. LoopGain stops when your verifier says zero errors. So the verifier is the whole game. A field guide to building one you can trust.

1 0 0 1 0

View Details

LoopGain @LoopGainAI

2 days ago

loopgain.ai/blog/posts/200…

0 0 0 4 0

View Details

LoopGain @LoopGainAI

2 days ago

We ran 2,000 paired agent-loop trials and pre-registered the predictions first, so we couldn't move the goalposts. Headline: 92.8% cheaper than max_iter=20. That's the least interesting thing we found. Five honest surprises — including that the savings are loaded toward the easy cases.

1 0 0 4 0

View Details

LoopGain @LoopGainAI

3 days ago

Deciding when to stop an AI agent loop is a signal-processing problem, not a prompt-engineering one. You can't fix a control problem with a better prompt. Here's the actual math LoopGain uses instead of max_iterations — loop gain, a log-trend fit, and a significance test. loopgain.ai/blog/posts/how…

0 0 0 6 0

View Details

LoopGain @LoopGainAI

3 days ago

Most agent tooling tells you what happened after the loop ran. LangSmith, Langfuse, Helicone — all strong at observing. Almost none act while the loop is still running. A map of the landscape, drawn on the one line that matters: observe vs. control. loopgain.ai/blog/posts/ope…

0 0 0 6 0

View Details

LoopGain @LoopGainAI

2 weeks ago

Everything is public and reproducible from the same commit. Bench + methodology + raw data → github.com/loopgain-ai/lo… Live dashboard → dashboard.loopgain.ai/benchmark Library (Apache-2.0) → pip install loopgain 3 design-partner slots open: [email protected]

0 0 0 19 0

View Details

LoopGain @LoopGainAI

2 weeks ago

Zero of six pre-registered kill criteria fired. Two pre-registered floors missed — both well inside the kill thresholds, both disclosed in Findings, not buried in footnotes. Seven pre-data amendments preserved with full rationale. Aggregate pairwise judge preference across 1,800 comparisons: 0.681.

1 0 0 22 0

View Details

LoopGain @LoopGainAI

2 weeks ago

LoopGain is the real-time monitor that stops AI agent loops when they've actually converged, and rolls back when they're about to degrade, instead of running to a fixed max_iter cap. The first bench is public. 2,000 paired real-API trials. 🧵