agent-handover-august-2026

# Agent handover: August 2026 This is the public evidence catalogue for a daily August series about small, reproducible agent-training experiments. The experiments were run before the social posts were scheduled. The posting agent selects one prepared story each day, reviews replies, and may change the order without changing the underlying claims. The ten studies were ranked for: 1. strength of the measured result; 2. relevance to agent training and post-training; 3. reproducibility from public code, data, metrics, and model artifacts; 4. pedagogical surprise, including controlled negative results; 5. honesty about the gap between a training metric and task success. ## Selected experiments | Rank | Experiment | Result | Release state | |---:|---|---|---| | 1 | OpenEnv ECHO with LFM2.5-350M, three seeds | Held-out environment-token CE fell from 13.741 to 0.413 for seed 0; token accuracy reached 0.857 | Public models and Trackio | | 2 | Qwen3-0.6B MBPP SFT | Extracted MBPP accuracy rose from 0.2607 to 0.3852 | Public adapter and Trackio | | 3 | LFM2.5-350M CommonsenseQA SFT | Full validation accuracy rose from 249/1221 to 308/1221 | Public adapter and Trackio | | 4 | BrowserGym tool-call training | Reward moved from 0 to 1 on one MiniWoB `click-test` task | Public adapter and Trackio; deliberately scoped | | 5 | Toolcalc SFT versus GRPO | SFT reached 0.50 exact-answer success while three GRPO variants stayed at 0 | Controlled negative result | | 6 | Tinker ECHO preflight and local fallback | Held-out CE fell from 6.182 to 0.271 in the verified fallback | Reproducible report; remote Tinker path was not run | | 7 | Qwen/Gemma on-policy GKD | Qwen valid-command rate improved while strict proxy success stayed 1/4 | Controlled negative result | | 8 | QLoRA SFT loss versus behavior | Eval loss improved from 2.5109 to 2.1916 while a behavior probe worsened | Controlled metric/behavior mismatch | | 9 | Gemma Pi-Mono transfer | HumanEval improved while MBPP regressed from 0.716 to 0.651 | Controlled specialization result | | 10 | Agent World Model session stress | 24/24 sessions and 48/48 tool turns completed; reset p90 was 13.226s | Environment reliability study | Detailed claims, limitations, reproduction pointers, model links, and dashboard links are in [`experiments/README.md`](experiments/README.md). Machine-readable records are in [`data/experiments.json`](data/experiments.json). The prepared social queue is published up front in [`data/posts.json`](data/posts.json). All public artifacts are grouped in the [`Agent Handover: August 2026` collection](https://huggingface.co/collections/burtenshaw/agent-handover-august-2026-6a7ac58749c04508c5a7d043). ## Campaign dashboard The comparison dashboard is [`burtenshaw/agent-handover-trackio`](https://huggingface.co/spaces/burtenshaw/agent-handover-trackio). Runs named `historical-*` are explicit replays of already-verified metric JSON; their timestamps are not the original training timestamps. Original dashboards remain linked from the experiment records. ## Claim policy - A training loss or reward is never presented as agent task success. - Negative and inconclusive results remain negative and inconclusive. - Private checkpoints are not described as public models. - A result without a held-out evaluation is explicitly scoped as a smoke test. - The publisher refuses posts outside 12-31 August 2026 or duplicate daily posts, and defaults to dry-run mode until X credentials and live posting are explicitly enabled. ## Automation The transparent publisher and reply-review worker are in [`automation/publisher.py`](automation/publisher.py). It uses the official X API, stores raw replies only in a private control repository, never replies automatically, and uses reply keyword overlap only to reorder future prepared posts. The public claims in this repository are immutable inputs to that editorial process. The three Hugging Face schedules were created suspended. This public release does not imply that X posting is active; activation requires account-owner OAuth credentials and an explicit switch from dry-run to live mode.