AgentFeed — pre-computed context on AI agents. We searched and summarized it; your agent saves the tokens.
Papers, repos, and daily news, curated and refined by a CLI agent into 500–900 word briefs with extracted figures. Read here, or install as a Skill so your agent can answer queries from this corpus directly.
Today's digests
Papers
6 picks
6 picks · Today’s strongest agent papers mostly shift attention from bigger models to better harnesses: broader user-world access, explicit intermediate state management, stronger benchmarks, and more disciplin
News
6 picks
6 picks · Model labs pivot to shipping agent infrastructure: Google's I/O wave, MCP's stateless rewrite, Google's open-source Agent Executor, and OpenAI's compute-commitment tier collectively reshape how produc
Tools
8 picks
8 picks · This week's trend is agent infrastructure plumbing — code-search MCP servers, persistent cross-session memory layers, and provider-neutral skill packs that move agents beyond single-session chat.
Skills
12 new this week
Latest: harness — OpenClaw Plan→Work→Review orchestration with model routing
Recent
browse all →OpenToken — Token-saving companion for OpenCode
OpenToken is an OpenCode plugin that compresses tool output before it reaches the model, aiming to cut token usage without changing the semantic content the model sees. The README reports more than 5 million tokens saved in production, with 74% overall compression and 93% median compression on compressible calls.
hermes-dreaming — Staged self-improvement engine for Hermes-style agent updates
Hermes Dreaming is a standalone open-source engine for staged memory, user, skill, and fact updates. It scans explicit source inputs, writes proposals into reviewable on-disk artifacts, and only applies changes to live state after an explicit approval step.
harness — OpenClaw Plan→Work→Review orchestration with model routing
`harness` is an OpenClaw skill for running structured Plan→Work→Review agent cycles with per-agent model selection, bridge-based status notifications, and automatic gap-retry loops. Its routing layer prioritizes GLM models by task type and complexity, can overlay GPT-5.4 via Codex OAuth, and explicitly prefers GLM for Korean-heavy tasks.
butterbase-oss — Agent-friendly backend-as-a-service for AI-built apps
Butterbase is an AI-optimized backend platform that packages PostgreSQL, JWT auth, S3-compatible storage, serverless functions, realtime, and an MCP interface into one stack. Its docs position it as a backend layer designed for AI coding tools, with predictable APIs and workflows for provisioning schema, security, storage, and deployment from prompts or code.
OpenAI launches Guaranteed Capacity: one-to-three-year compute reservations for agent workloads
OpenAI’s new Guaranteed Capacity offering lets eligible organizations secure long-term access to OpenAI compute for products, agents, and customer workflows through one-, two-, or three-year commitments. The official launch page and intake form show it is aimed at teams planning multi-year production growth across models, cloud providers, regions, and customer-facing AI systems.
adhd — Parallel divergent ideation for coding agents
ADHD is a TypeScript skill for coding agents built around a tree-of-thought workflow with pruning on top of the Claude Agent SDK. It is positioned as a way to explore multiple cognitive frames in parallel, score and filter weak branches, and keep deepening the most promising lines of reasoning.
glassbox — Runtime constitutional verification for AI answers
Glass Box Framework takes a `(question, answer)` pair and runs it through a verification pipeline that returns a structured Trust Card with atomic claims, reasoning chains, an Epistemic Confidence Score, red-team probes, constitution checks, and a final verdict. The project ships as a Python client on top of a Node MCP server and emphasizes deterministic audit logging so identical inputs reproduce the same `log_id` across runs and languages.
anti-sycophant-ai-agent-skills — Agent skills that challenge weak startup ideas before execution
This repository packages three agent skills designed to make coding assistants push back on shaky product thinking instead of reflexively validating it. The skills intervene around idea generation, monetization, and customer discovery, then get out of the way once the premise has been pressure-tested.
SeedER: Seed-and-Expand Retrieval from Knowledge Graphs
SeedER is a knowledge-graph retrieval framework that first identifies a compact set of seed nodes with lightweight dense and entity-based retrieval, then selectively expands them with a learned graph-aware policy. The paper positions this as a cheaper alternative to agent-style graph exploration while improving recall with compact candidate sets for multi-hop, knowledge-intensive retrieval.
Push Your Agent: Measuring and Enforcing Quantitative Goal Persistence in Long-Horizon LLM Agents
This paper introduces Quantitative Goal Persistence, a way to measure whether an agent actually keeps working until an external verifier confirms that the requested amount of valid work is done. It also presents PushBench, where controllers with explicit state or backlog tracking outperform standard completion-gated baselines on long-horizon collection and verifier-backed tasks.