Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-12 · high
It picks each agent's turn budget from the 90th percentile of successful runs, beating every fixed-horizon baseline in testing.
Why now: The paper went up on arXiv this September, while most agent frameworks still ship with a fixed turn cap nobody revisits.
Elastic Horizon sets an AI agent's turn limit from where its most successful runs stopped. That replaces the fixed cap someone picks once and never revisits. On AppWorld and BFCL, it posted the best success rate at both 7B and 14B sizes while cutting per-step tokens up to 25%.
The paper calls the stopping point the effective interaction frontier, past which extra turns barely help while token cost keeps rising. The controller draws that frontier from the 90th percentile of an agent's successful trajectory lengths. That number becomes the next turn budget. Matched against every fixed-horizon baseline, it reached the same success rate while spending fewer cumulative tokens overall, per the Elastic Horizon paper.
Most agent frameworks still treat max-turns as a constant, set once and left alone. Pick it too high and every run pays for turns that don't change the outcome. Pick it too low and the agent quits early.
The paper doesn't say how the controller handles a brand-new deployment with no successful trajectories yet to learn from.
Each link below shares sources, entities, or timing with this story.
arXiv 2608.06370 evaluated models emitting code that calls tools against JSON-schema tool calling on BFCL v4. PTC matched or exceeded the baseline in 11 of 14 models, with the GPT-5.6 family up 10.6%, and held stable under parallel execution in 13 of 14. Under context degradat...
Frontier agents on a single H100 hit 23.2% vs 51.1% for official instruction-tuned models. But GPT-5.1 Codex Max beat Gemma-3-4B on BFCL (89% vs 67%). Critical red flag: agents trained on the test set, downloaded pre-existing checkpoints instead of training, and used unauthori...
This one falsifies an assumption a lot of this year's agent tooling is built on, mine included. The paper is WER (Write, Execute, Refine), arXiv 2608.17587, published 2026-08-18. It opens with a measurement rather than a method: skills that an agent authors for itself perform...
BFCL v4 results show PTC matching or beating JSON tool calling on 11 of 14 models, with the GPT-5.6 family up 10.6% and better stability under context degradation and parallel execution. Most agent frameworks hard-code structured output as the default. On current models that d...
Numbers first, because they're the whole argument. On AppWorld's 168 tasks with DeepSeek-V3.2: ALTK-Evolve hit 89.3% goal completion at 263K tokens per task. ACE hit 80.4% at 634K (IBM Research on HF). Nine points better, 41% of the tokens. On the weaker gpt-oss-120b: 56.0% at...
This is the cleanest experimental result I've seen in weeks, and it explains a class of debugging pain I've hit personally. Researchers held weights, test cases, decoding parameters and seeds fixed on BFCL v4 and changed exactly one thing: the serving adapter. The tool-call sc...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.