Sol 5.6 Ultra gives surprisingly similar outputs every time, to the point where I've been investigating whether there's some illicit memory use thing going on... Don't think it's secret memory, just a bit RL fried + multiagent stuff collapses the randomness a bit (?)

In 2027, will most internet traffic be agents talking to agents? If yes — who wins: 1⃣the best models 2⃣the best infrastructure 3⃣the best governance? SVP Chain purpose-built for the agentic economy are positioning to capture all three. #AIAgents# #AgenticAI# #AgenticEconomy# #SVPChain# #MultiAgent# #FutureOfInternet#

Are your subagents re-reading the same files the supervisor already read? LangChain just shipped a feature to fix exactly that waste. Title: Organizing Context in a Multi-Agent Harness URL: 📝 Overview The deepagents framework now supports "forked subagents" — you can choose whether a subagent inherits the supervisor's full conversation history or starts from a blank context. ❗ Problem it solves Fully isolated subagents had to redo investigation and context-gathering the supervisor already completed, wasting both tokens and latency. ⚙️ Methodology Two modes are offered: Isolated Mode (fresh context) and Fork Mode (inherits the supervisor's full state). Fork Mode stays cost-efficient thanks to prompt caching. 🔧 Use cases ・Worker agents: use fork to continue fixing work the supervisor started ・Reviewer agents: use isolated for an unbiased evaluation ・Researcher and memory agents also pick modes based on their role 📊 Results No hard benchmark numbers are given, but the post reports reduced duplicate context-gathering and tool calls. I think this is a genuinely new lever for designing subagent roles well. #MultiAgent# #LangChain#

Packed house last night for our Omnigent & Multi-Agent Coding Meetup at the @databricks NYC office! 🗽 Highlights from the night: • Omnigent Overview • Training Custom Agents for Enterprise Tasks • Panel Q&A with @SpaceXAI, @Modal, and @Databricks A big thank you to our speakers and everyone who came out! 👏 🔗 Stay in the loop on future events: #Omnigent# #OpenSource# #NYC# #AI# #MultiAgent#

You can build any AI agent with any LLM. But when multiple agents start talking to each other, collisions are inevitable — because nothing is watching them. OpenAI agents may built them in a secret message board but it is still unmonitored. Agents only going to get wilder. That’s why putting agents on-chain with decentralized permission controls like SVPchain is what actually brings governance. #AIAgents# #MultiAgent# #AgenticAI# #OnChainAgents# #SVPchain# #AgentCollusion#

🤔 Does the communication topology for LLM multi-agent systems really need to be generated from scratch every single time? A team at UCLA questioned that very assumption. Title: Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems URL: ❓ Why have people assumed agent communication topologies need to be searched by a generative model every time? 💡 Turns out they may not need to be. The topologies that survive reward filtering converge to only about 6 distinct graphs, even as the codebook size grows from 8 to 64 — the design space is far smaller than it looks. ❓ Wouldn't making the graph sparser (fewer edges) cut down token usage? 💡 The opposite happens. Edge count correlates with token cost at r=-0.4 — sparser graphs actually use more tokens. Structural "cheapness" doesn't match measured cost. ❓ What's wrong with scoring candidate topologies using a GNN, as prior work does? 💡 For homogeneous agent teams (common in real deployments), the GNN's message passing treats every adjacency matrix as the same input, so it can't tell candidates apart — a real failure mode, not just a minor gap. ❓ So how well does Codebook Agent actually work once it swaps generation for selection? 💡 By compressing topologies into 16 codes via a VQ-AE and picking candidates with a reward-weighted MLP plus a proxy trained on real measured costs, it cuts generation latency from 301-396ms down to 2.4ms — a 125-158x speedup — while improving accuracy by 1.6 points on average over the prior best method and cutting tokens by 21.9-33.2%. #MultiAgent# #LLM#

Multi-agent collaboration that exchanges text every round is expensive 🔄 The fresh idea: scale the collaboration itself through latent-space recursion. Title: Recursive Multi-Agent Systems URL: 🔄 Overview RecursiveMAS reframes multi-agent collaboration as recursive computation in a unified latent space, rather than sequential text exchange. It connects heterogeneous agents through a RecursiveLink module, enabling latent thought generation and cross-agent state transfer. ❓ Challenges Solved Multi-agent systems (MAS) typically rely on text-based communication. ・Agents conversing in natural language consume huge token counts and cost a lot of compute ・The driving question: can agent collaboration itself be scaled through recursion? 💡 Methodology & Proposed Approach ・It frames the whole system as recursive latent-space computation ・The RecursiveLink module lightly connects heterogeneous agents with gradient-based credit assignment across recursion rounds ・Optimization uses inner-outer loop learning while maintaining theoretical stability 📊 Experimental Results Across 9 benchmarks spanning math, science, medicine, search, and code generation, with 4 collaboration patterns tested. ・Average accuracy improved by 8.3% ・Inference sped up 1.2x-2.4x ・Token usage reduced by 34.6%-75.6% It raises accuracy while substantially improving speed and cost at the same time. #MultiAgent# #LLM#