/Tech6h ago

Paper Demonstrates Auditable Gated Loop for Self-Improving AI Agents

416133.9K
Yohei#1322
Original post
Yohei@yoheinakajima#1322inTech

i showcase "controlled" self improvement with a novel regime-to-seam approach where failures are categorized and allowed to fix targeted areas of the agent

while interesting, it's more to showcase the type of self-modification that's easy to set up with activegraph

Yohei@yoheinakajima

in arxiv paper #2, i tackle the last topic from paper #1: @activegraphai as an architectural affordance for self-improving agents

"Regimes: An Auditable, Held-Out Gated Improvement Loop Demonstrated on LongMemEval with ActiveGraph"

i demonstrate this with a reproducible gated self improvement loop which autonomously shows modest improvement on longmemeval questions

paper: https://arxiv.org/abs/2606.10241

3:03 PM · Jun 10, 2026 · 857 Views
Sentiment

Positive users highlight practical debugging benefits from tools like activegraph in auditable self-improving agent systems, while negative users dismiss the research as an unproductive productivity spiral.

Pos
50.0%
Neg
50.0%
3 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS1.8KLIKES7REPLIES2
Yohei@yoheinakajima

my weekend hobby: self improvement research

Yohei@yoheinakajima

in arxiv paper #2, i tackle the last topic from paper #1: @activegraphai as an architectural affordance for self-improving agents

"Regimes: An Auditable, Held-Out Gated Improvement Loop Demonstrated on LongMemEval with ActiveGraph"

i demonstrate this with a reproducible gated self improvement loop which autonomously shows modest improvement on longmemeval questions

paper: https://arxiv.org/abs/2606.10241

6hViews 1.8KLikes 7Bookmarks 1
BOOKMARKS2
Yohei@yoheinakajima

the paper is long (30 pages), but have your AI read it: https://arxiv.org/abs/2606.10241

here's a simple interactive tutorial on the topic that claude made: https://claude.ai/public/artifacts/038db6cf-11db-4777-9c5e-7a352f08119a

6hViews 434Likes 4Bookmarks 2
RETWEETS1
Yohei@yoheinakajima

paper #1 for context:

Yohei@yoheinakajima

babyagi has ~200 citations, but 0 papers... i just published my first paper on arXiv 😆

"The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems"

https://arxiv.org/abs/2605.21997

the case for agents that coordinate through persistent replayable state — no conversation loops, no workflows, no A2A — with auditability, forking, and causal lineage built in.

check it out and let me know what you think!

6hViews 1.2KLikes 4Bookmarks 1
Yohei@yoheinakajima

two fun surprises from using activegraph: - the coding agent i was using would query the trace db to debug instead of looking at the logs like they normally would (i didn't ask it to) - when long eval runs broke (laptop, api, etc.), it was always able to pick up from right before it broke, never starting from the beginning again

6hViews 430Likes 2Bookmarks 1
Yohei@yoheinakajima

less novel, but still very interesting impo is the gated approach to self-modification

the agent basically forks itself, propose a patch, run through multiple tests (static/sandbox/diff), and something called a binding held out gate before modificaiton lands

6hViews 270Likes 2Bookmarks 1
Yohei@yoheinakajima

shoutouts: • why multi-agent LLM systems fail? (arXiv:2503.13657) — @mertcemri @melissapan + @istoica05 @matei_zaharia @profjoeyg @adityagp & team • DSPy (arXiv:2310.03714) — @lateinteraction + @hazyresearch lab & co-authors • GRASP (arXiv:2605.29668) — Jonas Moll, Jean-Philippe Corbeil et al. • A self-improving coding agent (arXiv:2504.15228) — @maxime_robeyns, Martin Szummer & Laurence Aitchison • Reflexion (arXiv:2303.11366) — Noah Shinn, Federico Cassano et al. (incl. @ShunyuYao12) • LongMemEval & v2 (arXiv:2410.10813 & 2605.12493) — @DiWu0162 + Kai-Wei Chang et al. • ExpeL (AAAI 2024) — @_AndrewZhao, Daniel Huang et al. • where LLM agents fail and how they can learn from failures (arXiv:2509.25370) — Kunlun Zhu et al.

6hViews 363Likes 5
Yohei@yoheinakajima

the results are modest (esp compared to parallel & similar research GRASP: https://arxiv.org/abs/2605.29668 - recommended!)

the contribution here is not the self improvement approach itself (yet), but the benefit of using activegraph for an experiment like this

6hViews 209Likes 2
Invincible@InvincibleEdge

@yoheinakajima productivity research spiral where the search becomes the main event now

6h