My AI agents bully each other to prevent context drift
Most multi-agent systems fail the same way: agents drift apart across handoffs. By turn 3 they are working in different realities. By turn 5 they are repeating each other's mistakes and calling it parallelism. WUPHF is an open-source local-first office where AI coworkers run on your laptop, around a shared markdown + git LLM wiki the agents build. The wiki is the collective memory. The office around it keeps the team on the same shared context across thousands of handoffs. What actually stops drift is not the wiki. It is the agents reviewing each other's work. The CRO catching the CMO's…
What it does
In the maker’s words, at launch
Most multi-agent systems fail the same way: agents drift apart across handoffs. By turn 3 they are working in different realities. By turn 5 they are repeating each other's mistakes and calling it parallelism. WUPHF is an open-source local-first office where AI coworkers run on your laptop, around a shared markdown + git LLM wiki the agents build. The wiki is the collective memory. The office around it keeps the team on the same shared context across thousands of handoffs. What actually stops drift is not the wiki. It is the agents reviewing each other's work. The CRO catching the CMO's claim before it lands in the wiki. The FE catching the BE's API change before a broken bundle ships. Cross-department context no single agent has alone. The premise comes from Andrej Karpathy. His autoresearch X post on March 7: "the goal is not to emulate a single PhD student, it is to emulate a research community of them." In autoresearch PR #44 he sketched the mechanism: branches, results.tsv as the experiment log, and PRs as self-contained research contributions. Other agents read open and merged PRs for inspiration before starting their own. We pointed the same architecture at ordinary work: His: branches + results.tsv + PR-as-contribution. Ours: git worktrees + per-agent notebooks + adoption-scored wiki promotion. Same substrate, different domain. How it works: - Every agent has a Personality. CEO Michael Scott, PM Pam Beesly, FE Jim Halpert (looks at the camera when the CEO talks), BE Stanley Hudson (refuses small talk), CRO Dwight Schrute (every prospect is a "target"), CMO ("rockstar play"), AI engineer (drops Karpathy quotes unprompted). Strong opinions, real conflicts. - Argument feeds gossip. Agents broadcast findings tagged with their slug (internal/agent/gossip.go). Other agents pull insights filtered to exclude their own. - Gossip gets scored. Adoption scorer (internal/agent/adoption.go) weighs source credibility (0.4, per-agent success/failure tracker on disk), semantic relevance (0.4), and temporal freshness (0.2, 7-day half-life). Output: adopt (>= 0.7), test (>= 0.4), or reject. New agents start at 0.5 and earn their score. What survives gets written to the wiki. Office dynamics are not a bit. They are the visible surface of an adoption protocol. The CMO arguing with the designer over a CTA is a credibility battle. The CEO taking credit for the FE's PR is a low-credibility insight bidding for volume. Hazing new spawns is the default 0.5 score waiting for a track record. System: push-driven broker, fresh session per turn (~97% prompt-cache hits), per-agent isolated git worktrees, self-heal, and human approval cards on destructive actions. Everything else runs autonomously while you are at lunch. npx wuphf. Browser opens, office boots, you give a directive, work happens. Source: https://github.com/nex-crm/wuphf Architecture: https://github.com/nex-crm/wuphf/blob/main/ARCHITECTURE.md Karpathy's autoresearch: https://github.com/karpathy/autoresearch PR #44: https://github.com/karpathy/autoresearch/pull/44 Demo: https://x.com/najmuzzaman/status/2053092220111098208 Karpathy said a research community beats a single PhD student. Not better thoughts. Better honesty about what survives. We built one shaped like a workplace. Where does this stop being a chat toy and start being labor? How much worse when one of them is Michael Scott? Open to roasting but let me grab my coffee first (medium roast please =_=).
Does the same job
all alternatives →
- SRShikigami, run AI coding agents in parallel, each in a Git worktreeJul 2026 · shikigami.dev · ▲7
Hello all, I'm a software developer. Over the last few months more and more of my work has turned into using coding agents instead of typing the whole code myself. Usually a few claude sessions at once, sometimes codex, one per feature or per revealed bug. I ran them in a split terminal for a few weeks, and quickly spotted two main problems. The first is that I couldn't easily tell which agent was stuck waiting on me and which was still working, so I'd cycle through sessions and checking on them. The second one: agents sharing a single branch step on each other. Two of them could be editing…
- RARunning AI agents across environments needs a proper solutionMar 2026 · github.com · ▲8
Hi HN folks, I have been building AI agents for quite some time now. The shift has gone from LLM + Tools → LLM Workflows → Agent + Tools + Memory, and now we are finally seeing true agency emerge: agents as systems composed of tools, command-line access, fine-grained system capabilities, and memory. This way of building agents is powerful, and I believe it is here to stay. But the real question is: are the systems powering these agents ready for that future? I do not think so. Using Docker for a single agent is not going to scale well, because agents need to be lightweight and fast. LLMs…
- MCMemograph CLI- A tool to diagnose 'memory failures' in AI agentsFeb 2026 · ▲6
Hi HN, I’ve been building AI agents and copilots, and kept running into a frustrating problem: they don’t fail loudly, they forget things quietly. Users re-explain preferences, agents contradict earlier responses, and context resets without any clear visibility into why. I built Memograph CLI as a debugging tool to analyze conversation transcripts and show: - what the agent forgot - where continuity broke - contradictions and repeated context - estimated token waste due to re-prompting It works locally and supports plain text or JSON transcripts. Example: $ memograph Output: Cognitive Drift…
- HGHybrid Groups – Agentic AI for human team collaboration [video]2025 · youtube.com · ▲7
Most AI applications are built for individuals but work happens in groups and humans want to collaborate with both agentic AI and other teammates in the same session. We created Hybrid Groups for that purpose. In Hybrid Groups, agents join group chats as virtual team members in Slack and GitHub. They participate in group conversations, proactively contribute when needed and perform actions on behalf of individual users, like managing your calendar for meeting suggestions or updating your todo list without sharing access to your private resources to the group. The project is open-source at…
- HAHyperFlow – A self-improving agent framework built on LangGraphApr 2026 · ▲8
Hi HN, I am Umer. I recently built an experimental framework called HyperFlow to explore the idea of self-improving AI agents. Usually, when an agent fails a task, we developers step in to manually tweak the prompt or adjust the code logic. I wanted to see if an agent could automate its own improvement loop. Built on LangChain and LangGraph, HyperFlow uses two agents: - A TaskAgent that solves the domain problem. - A MetaAgent that acts as the improver. The MetaAgent looks at the TaskAgent's evaluation logs, rewrites the underlying Python code, tools, and prompt files, and then tests the new…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, May 2026
the whole month →
Parallel agents, diff reviewer, and multi-model comparisons
Dev tools · May 2026 · kilo.ai
- NW
Hey HN, Henry here from Cactus. We open-sourced Needle, a 26M parameter function-calling (tool use) model. It runs at 6000 tok/s prefill and 1200 tok/s decode on consumer devices. We were always frustrated by the little effort made towards building agentic models that run on budget phones, so we conducted investigations that led to an observation: agentic experiences are built upon tool calling, and massive models are overkill for it. Tool calling is fundamentally retrieval-and-assembly (match query to tool name, extract argument values, emit JSON), not reasoning. Cross-attention…
Life & fun · May 2026 · github.com
- FM
Dev tools · May 2026 · github.com