138 links · newest first
This Turing Post video examines DeepSeek's open-sourced 'Harness,' a coding-agent architecture built around the principle that everything is a plugin, capable of writing and removing its own tools mid-task via the Cordis runtime. It connects the design to historical shifts like Smalltalk, Unix, and…
Visit originalAGNTCon + MCPCon North America is a Linux Foundation conference taking place October 22-23, 2026 at the San Jose McEnery Convention Center. It brings together developers, researchers, platform builders, and enterprises advancing AI agents and MCP, including agent architectures, infrastructure, prot…
Visit originalA comprehensive survey (arXiv:2504.01990) proposing a brain-inspired modular framework for foundation agents, covering memory, world models, emotion, and reward systems as analogues to biological cognition. It surveys self-evolving agents that improve via LLM-based optimization and AutoML, multi-ag…
Visit originalThe video breaks down DeepSeek Harness's append-only, event-sourced session log and invariant checker, explaining how byte-for-byte request reconstruction keeps prefix caching stable and cached tokens roughly 120x cheaper than uncached ones. It covers the Cordis plugin meta-framework, four operatin…
Visit originalThe page announces "Harness Engineering: Design Patterns for Securing Long-Horizon Multi-Agent AI Systems," a Kindle book about transitioning from software engineering toward harness engineering for non-deterministic agentic AI systems. It presents ten pattern families covering autonomy, identity, …
Visit originalThe video explains how graph engineering extends loop engineering by splitting tasks across parallel AI-agent nodes connected by data-flow edges, using structures like diamonds and fan-in barriers. It covers verification methods including standalone, embedded, second-opinion, and orchestrated skill…
Visit originalThe post argues that running 300 agents in a swarm to finish research quickly just produces a pile of separate documents unless their findings are linked together. It describes connecting the outputs of 300 agents into a single knowledge graph, representing the agents as nodes and their relationshi…
Visit originalLight Society is an agent-based simulation framework that uses LLM-powered operations, distilled surrogate models, event aggregation, and compressed networks to simulate societies with over one billion agents. Using World Values Survey demographic profiles, the paper evaluates Trust Games, opinion …
Visit originalThis first article in a four-part series examines how AI agents can use short-, medium-, and long-term memory across users, agents, and Kubernetes clusters. It defines memory types, compares baseline and production memory implementations, and surveys dedicated frameworks, native agent-framework mem…
Visit originalStanford's Virtual Biotech organizes tens of thousands of specialized AI agents across divisions such as target discovery, molecule design, and clinical trials. Its Paperclip platform builds an AI-native virtual file system that helps agents process scientific data more accurately while cutting tim…
Visit originalThe post presents memory engineering as the deliberate design of what an AI agent stores, retrieves, prunes, and forgets, drawing on perspectives from Stanford, Microsoft, Anthropic, and Nvidia. It covers write-path energy costs, fact-and-skill extraction, file-based control, KV-cache limitations, …
Visit originalThe post explains how Claude can progress from basic chat to autonomous agents by using tools, memory, and iterative workflows. It describes self-checking loops with measurable criteria and stop conditions, then introduces graphs that parallelize independent tasks through fan-out and convergence. I…
Visit originalContent behind a login wall / unavailable — inferred from the link preview: a post by @cyrilXBT quoting Andrew Ng saying AI agents are now doing almost 100% of his tasks, exceeding his expectations, and predicting that within 3-6 months people will be building graphs to orchestrate self-improving a…
Visit originalBerkeley RDI and the Responsible Decentralized Intelligence Foundation hosted the Agentic AI Summit on August 1-2, 2026, at UC Berkeley in Berkeley, California. The two-day event covered Agentic AI research, infrastructure, industry applications, evaluation, safety, security, and deployment, with a…
Visit originalThe video examines Tencent Cloud's open-source memory plugin for the OpenClaw agent runtime, which reduces context rot by compressing intermediate execution logs into Mermaid graphs and storing detailed evidence locally. It explains the plugin's four-layer memory architecture, SQLite-vec hybrid ret…
Visit originalThe article explains how recent model releases like GPT-5.6 Sol, Kimi K3, and Claude Opus 5 enable AI agents to sustain multi-step work rather than just produce single responses. It argues effective multi-agent systems require deliberate graph engineering, with clear roles, handoffs, review loops, …
Visit originalThis Substack article explains Moonshot AI's Kimi agent swarm system, focusing on how centrally orchestrated multi-agent setups tackle long-horizon tasks, illustrated by a gravitational-wave analysis example. It compares Kimi's design to classical swarm intelligence seen in ants and bees, and cover…
Visit originalThis X post argues that manually re-prompting AI models like Claude is an outdated workflow, advocating instead for building autonomous 'loops' that let models self-correct and retain memory across sessions. It references an Anthropic engineer's advice to build the loop once and let it do repeated …
Visit originalThis page could not be fetched directly due to a site-side access restriction, but based on the URL structure from The New Stack (a developer/technical publication), it appears to discuss portability of AI agents across platforms and the need for a standardized 'contract' or interface so agents bui…
Visit originalThe content presents a decision framework for whether a task calls for a chat, a single agent, a team of agents, or no AI assistance at all. It evaluates tasks using four estimates — size, independence, separation, and checkability — and explains how spending, verification, and context limits affec…
Visit original