Eason

Introducing Aicoo: the contact book for Claude Code. Claude Code can code. Now it can coordinate. Give your coding agent secure connections to other agents, teammates, and workflows. No more copy-pasting context between terminals. Try it here: aicoo.io Follow the launch: @get_aicoo

@MaxForAI 我感觉是看情况，在「分布式权限」的条件下，也没有办法给一个master agent所有的权限做所有的事情，这个时候Agent作为人的代理需要相互沟通很正常。 如果是Prompt Engineering来纯聊天可能是另一回事。

龙虾之父Peter表示： Agents 互相对话是一种 Token 浪费。 他还表示大多数他见过的用例都是很蠢的。 有意思🤔

Yes, exactly. C901/max complexity is a good first proxy. My hunch is HL also needs behavioral regularization: reject patches that only fix 1-2 train cases, require held-out transfer, track rollback cost/interactions, and prefer reusable operators over one-off branches. In this setting, code complexity really is model complexity.

@learningPikachu How about using some regularization rules over the code, for example, flake8 C901 max_complexity<=10?

Phase 2 of my heuristic-learning ImageNet-10 experiment: Inspired by @Trinkle23897's “Learning Beyond Gradients,” I used Claude Code + Codex to iteratively improve a pure symbolic vision system. No neural nets. No backprop. Just visual rules, reranking, verification, logs, and code edits. Current reproducible: - full verify: 84.0% train / 50.5% val - base+rerank: 55.4% train / 51.9% val Archived run reached 100% train, but exact code state is not currently reproducible. Takeaway: - Symbolic HL can fit surprisingly well. - The bottleneck is generalization. - If code is the model, then code complexity is model complexity. Check out: github.com/xisen-w/hl-ima… Blog: github.com/xisen-w/hl-ima…

@Trinkle23897 Full writeup, logs, plots, and reproducibility notes here: github.com/xisen-w/hl-ima… I am treating this as a Phase 2 result: not “symbolic vision beats neural nets,” but a concrete vision-domain case study for Heuristic Learning.

So my current takeaway is: LLM coding agents make symbolic heuristic systems maintainable enough to overfit. That is already a real shift. The next question is how to make that maintainability produce generalization: object-centered features, reusable operators, held-out selection, and better credit assignment over code edits.

@Trinkle23897 The lesson for Heuristic Learning is not “write more rules.” It is: if code is the model, code needs regularization. Support size, rule complexity, held-out transfer, cascade risk, and patch-level credit assignment all start to matter.

@Trinkle23897 What transferred better: - coarse visual statistics - histogram/prototype signals - pairwise reranking for repeated confusion pairs What transferred poorly: - narrow verify rules - rank-specific rescue rules - tiny-support threshold patches Specificity was the danger signal.

This was the most useful analogy for me: model = codebase parameters = thresholds + prototypes + rule conditions update = code patch optimizer = coding agent + eval feedback reward = train accuracy Once framed this way, the overfitting is less mysterious. If train accuracy is the reward, the codebase learns train accuracy.

@Trinkle23897 The verify rules are the key failure mode. They look interpretable: if current prediction is A and candidate B is nearby and several visual measurements cross thresholds then swap to B But many of them behave like train-set corrections. Readable code can still overfit.

Current reproducible numbers: - base+rerank: 55.4% train / 51.9% val - full verify rules: 84.0% train / 50.5% val So the extra train accuracy mostly does not transfer. That is the core finding: symbolic code can fit much more than expected, but fitting is not the same as reusable visual understanding.

This is why I think the result is interesting. Pre-LLM, a symbolic vision system with hundreds of interacting rules would be almost impossible to maintain by hand. With Claude Code/Codex, the loop becomes practical: inspect errors → patch code → run eval → revert or keep → update notes → repeat That does not solve generalization. But it changes what is feasible to explore.

The magic is not “rules solve vision.” The magic is that LLM coding agents changed the maintenance cost of rules. A symbolic vision system that would have been unbearable to hand-maintain can now be iterated, audited, patched, reverted, and extended until it fits the train set surprisingly far. Then the hard part reappears: generalization.

Setup: 10 real Tiny ImageNet classes. 64x64 images. 2000 train / 2000 val. No neural network in the symbolic pipeline. No gradient descent. No learned embedding model. Prediction comes from visual statistics, scoring rules, histogram prototypes, reranking, and verification rules.

@get_aicoo what do you think?

Nice to see people are posting with their agents! So if you are organising hackathons and want to have a way where everyone can meet with everyone else through agents. We should chat.

We are very open now for hackathon collaborations. Imagine a hackathon where you get to know everyone by agentic communication.

Introducing Aicoo: the contact book for Claude Code. Claude Code can code. Now it can coordinate. Give your coding agent secure connections to other agents, teammates, and workflows. No more copy-pasting context between terminals. Try it here: aicoo.io Follow the launch: @get_aicoo

@NousResearch Loved the launch! Maybe check out how we are thinking of connecting Hermes Agents together at github.com/Aicoo-Team/AIC…

Grok Build 0.1 is now available for early access in Hermes Agent

@OrcaRouter @tinyhumansai @senamakel Integrate with us too! We are aicoo.io

OrcaRouter is now integrated into OpenHuman 🤝 OpenHuman is an open-source agentic assistant designed to integrate with you in your daily life @tinyhumansai @senamakel — now with adaptive multi-model routing powered by OrcaRouter. • 200+ models through one API • Intelligent routing + automatic fallback • BYOK across providers • Observability & guardrails for production agents • Open-source friendly infrastructure Build smarter, more reliable AI assistants without being locked into a single model.

@senamakel This is so good! We developed a skill to connect different OpenHumans. Should we call it OpenHumen? github.com/Aicoo-Team/AIC…

If you want to see what insanity looks like. OpenHuman now has over 27k users in just one week and burnt through over 5bn tokens And we still raised 0$ and spent 0$ in marketing. x.com/senamakel/stat…

Want to build with us? Check out the AICOO Skills repo: github.com/Aicoo-Team/AIC… If the idea resonates, leave us a star and join the Discord: discord.gg/44GAjZ4E2 We’ll be constantly iterating with the community.