Search posts, papers, and topics
This empirical study dissects coding agent harness components like planning, action space, and context management. It finds context management is critical for tight windows, rule-based elision is efficient, and planning shifts from accuracy to cost savings for stronger models.
MintAct is a family of vision-language models (2B-8B) that unifies UI grounding, multi-step navigation, and visual tool use across mobile, desktop, and web environments. It achieves state-of-the-art performance by leveraging a scalable environment and an asynchronous reinforcement learning infrastructure.
Production AI agents often fail silently due to issues beyond model hallucination, such as state inconsistencies. This talk outlines principles for building reliable agent harnesses, focusing on explicit state ownership, ordered mutations, and verifiable actions, using real-world case studies.
ScienceBuddy is an interactive workspace that converts researcher prompts, feedback, and execution traces into continual‑learning tasks for AI agents. It introduces a "recursive‑in‑recursive" self‑improvement loop that alternates harness refinement and model training, and showcases case studies across four scientific task families.
A step‑by‑step tutorial showing how to add a typed decision layer (using Jev and Pydantic) to a LangChain agent so you can route requests to cheaper or more capable models and gate risky tool calls, with concrete Python code and cost‑trade‑off notes.
Evaluating AI coding agents requires a robust sandbox to prevent agents from retrieving answers from the environment, which can invalidate tests of internal knowledge. A correct answer doesn't guarantee a valid measurement if the agent accessed information it shouldn't have, highlighting the need to define sandboxes by information boundaries rather than just tool restrictions. Always review agent…
The article shows how to build a deterministic, checkpoint‑backed finite state machine engine in TypeScript for orchestrating multi‑agent AI workflows. It uses Zod for schema validation, better‑sqlite3 for atomic persistence, and a pure transition function to make workflows traceable and recoverable in serverless environments.
Matt Pocock explains how he uses AI agents for software development, emphasizing "strategic programming" and guiding agents with "leading words" from classic engineering texts. He argues that this approach makes engineering fundamentals more critical than ever for creating agent-optimized codebases.
APort Vault is a benchmark that replays 4,371 human‑written attacks against a live payment‑handling AI agent across 14 models and multiple policy configurations, generating 225,964 evaluations. Adding the Open Agent Passport pre‑action check eliminated all unauthorized transfers in the test, showing a per‑session breach upper bound of 0.38%.
The author turned a one‑off script that rendered each Jekyll page to Markdown and added a `<link rel="alternate" type="text/markdown">` tag into a reusable Jekyll plugin (jekyll‑md). The gem hooks into the build, extracts the main content (defaulting to `<main>` or `[role="main"]`), converts it to Markdown, injects the discovery link, and requires no per‑page front‑matter. Configuration is option…
A study found that an LLM's stated knowledge cutoff date is a poor indicator of its actual capability regarding specific product versions. Models showed inconsistent knowledge, failing on older features while sometimes inferring correctly about newer, post-cutoff features.
Google Cloud adds Filestore agent volumes, a fully‑managed, elastic file‑system that automatically provisions isolated POSIX workspaces for GKE‑based AI agent sandboxes. Volumes attach in milliseconds, support RWX with file‑level locking, and charge only for used capacity with automatic tiering, aiming to cut cold‑start latency and storage waste for large‑scale agent fleets.
This podcast discusses the critical challenges of identity, authorization, and security for AI agents. It introduces the DPACT framework (Delegation, Policy, Auditability, Context, Time) as a blueprint for building responsible agentic systems with bounded, delegated authority, moving beyond simple token-based access.
Agent Substrate is an open‑source runtime for AI agents that runs on GKE. It uses Cloud Hypervisor microVMs or gVisor sandboxes to give kernel‑level isolation, a custom control‑ and data‑plane that can suspend/resume agents in <500 ms, and a “zero‑idle” model that packs >1 000 dormant agents per host (≈10× density vs. containers). GKE integration adds custom ComputeClasses, spot/on‑demand pools,…
Orange created a FinOps Community of Practice and runs ‘Clean Days’—protected, gamified sessions that get engineers actively optimizing cloud spend. They then scale this culture with AI agents that surface cost data, suggest quick‑win fixes, and automate reporting, extending responsibility across thousands of engineers.
Databricks outlines five criteria for a production‑ready database for AI agents—branch‑per‑agent isolation, serverless scale‑to‑zero, hybrid search, ACID guarantees, and a unified platform that eliminates ETL lag—illustrating each with features of its Lakebase offering and brief customer anecdotes.
Auth0’s early‑access FGA Permissions Index pre‑computes all implied user/agent permissions on write, streams the flat index into the customer’s own database, and lets applications enforce permissions with a simple SQL join instead of runtime graph traversals. The feature targets enterprise B2B SaaS workloads with high‑cardinality relations and AI‑driven agents, promising constant‑time lookups and…
Datadog explains how scaling AI agents turns isolated executions into shared‑platform risk and outlines a systematic approach to model agent trajectories, monitor per‑dependency constraints, and enforce workload‑specific capacity policies. The result is proactive detection of bottlenecks and protection against noisy‑neighbor failures.
