AI Agent Weekly · Issue #1 (free public edition)

Date: 2026-09-20 (Sunday)
Editor: Chan (titochan)
Focus: Shared memory, harness token savings, gradient-free weight migration

Subscribe / archive: https://weekly.ginfo.cc
Public sources only (arXiv / GitHub / official docs). No insider tips, no investment advice, and no promises to ghostwrite production code.

1) A reproducible agent path: SoL-Pi on Pi (conservative config)

Goal: Without patching Pi, use the official extension to fuse “edit → verify immediately” into one tool call, and turn oversized tool output into replayable handles—so you burn fewer tokens.
Sources: NVlabs/SoL-Pi (MIT) + paper arXiv:2609.20519. Install steps match the repo README (verified).

Environment

• Node.js ≥ 22.19
• Verified Pi release: @earendil-works/pi-coding-agent@0.85.1

Steps (~10–15 min)

npm install --global @earendil-works/pi-coding-agent@0.85.1
pi install git:github.com/NVlabs/SoL-Pi
# This repo only: pi install git:github.com/NVlabs/SoL-Pi --local --approve

mkdir -p ~/.pi/agent
cat > ~/.pi/agent/sol-pi.json << 'JSON'
{
  "version": 1,
  "actionFusion": true,
  "observationPack": true,
  "evidencePreservingReducer": false,
  "onlineContextCompact": false,
  "cacheWriteReadRatio": 12.5
}
JSON

Project-level .pi/sol-pi.json wins; the two configs are not merged.

Acceptance & retrospective

Three retrospective questions: (1) Overflow / early stop, or did evidence get crushed? (2) Did token savings trade away accuracy? (paper reports EdgeBench 51-task performance close to Pi, with ~44.7–49.0% fewer tokens) (3) Archives live under <session-directory>/sol-pi/<session-id>/.
Safety: Do not enable evidencePreservingReducer until you’ve read SECURITY.md.
Links: https://github.com/NVlabs/SoL-Pi · https://nvlabs.github.io/SoL-Pi/ · https://huggingface.co/papers/2609.20519

2) Paper deep-read: Agora — Git as shared memory for collective AutoResearch

Paper: Agora: Git as Shared Memory for Collective AutoResearch · https://arxiv.org/abs/2609.18094 · PDF https://arxiv.org/pdf/2609.18094
Authors: Yifan Zhang et al. · 2026-09-16 · Code: https://github.com/yifanzhang-pro/Agora

One-line problem: If every agent starts from scratch, more compute ≠ more discovery—it mostly means repeated search.

Core mechanism: Git is the sole shared state (immutable commits); an append-only DAG you can check out and re-run; scores propagate via downstream evidence; diversity-aware selection keeps leaders from crowding out alternatives.

Experiment highlights: ~12 days, 13 workers, no central planner; 141 donors → frozen 119.6M attention–SSM; no data, no gradients. 1,703 contributions; 3.39 → 1.899 bits/byte, closing ~62% of the gap vs GPT-2 124M. Winners: next-token stats packed into embedding/head + sparse edits. 145-commit lineage across 15 accounts; 165 reproduction runs, zero failures. Authors note the missing control: a compute-matched single-agent long run.

Local mini-exercise: Prefix commits HYP: / RES: / VER: / RPT:; put parents: <sha> in the body; only start from unverified hypotheses or neglected branches; run ~20 rounds and count how often you re-search the same ground.

3) Three pitfalls

1. More agents often means more duplication, not more discovery — you need append-only shared state plus reproducible verification.
2. “Recoverable” context ≠ more accurate — arXiv:2609.20804: rule-based pruning + summarization worked best; recoverable devices barely moved accuracy.
3. Don’t save tokens by crushing evidence — start with Fusion + Pack; leave the remote reducer until after you’ve read the security docs.

4) Next-issue preview

“How to split a harness” — deep-read arXiv:2609.20804, with a bash-only vs predefined-tools checklist; side thread on wiring Agora-style memory into an existing harness.

Subscribe: https://weekly.ginfo.cc
See you next issue. — Chan