Tag
Boris Cherny and others describe shifting from prompting AI agents to designing autonomous loops that run continuously, using memory files and evaluator patterns for code quality.
The article argues that current AI agent frameworks treat agents as black boxes, making them unmaintainable, and proposes a Git-native architecture (Lyzr GitAgent, OpenGAP) where agent logic is version-controlled as flat files with pull requests for rollback and auditability.
A developer describes the challenge of building multi-agent AI assistants that fail to handle unexpected situations gracefully, relying on explicit rules that lead to a whack-a-mole problem instead of enabling autonomous reasoning about ambiguity.
The author proposes a mental model where AI agents should maintain a separate memory layer (brain) that stores reusable understanding, distinct from their knowledge base (library), to avoid rediscovering the same information repeatedly.
The tweet recommends an article on agent architecture in production, highlighting the use of Traces to diagnose issues and implement an iterative improvement loop.
The article argues that an AI agent is defined by its durable event log, not the runtime or model, enabling fault-tolerant resumption and simplified reasoning about agent state.
Anthropic released Fable 5, a powerful new model with high pricing, making cost-aware routing essential for agent builders due to token fan-out and high output costs.
The article discusses the challenge of determining which AI agent actions require human approval, citing a $27M unauthorized transfer in January 2026, and proposes a framework based on reversibility and impact.
AdaCoM is a separate LLM that manages context for a frozen AI agent, improving performance on long tasks without retraining. It improved average web search performance by 39% in tests.
This paper introduces Queen-Bee, a governed multi-agent architecture for enterprise MCP orchestration that separates planning and execution via a BeeSpec intermediate representation, achieving high task success rates with zero governance failures in prototype evaluations.
A former Google CEO shares the key directions for learning AI (agents, Claude Code, prompts, etc.) and recommends a series of free official learning resources from LangChain and Anthropic.
A tweet describes the concept of gBrain as the company brain, an orchestrated system of AI agents organized into department verticals with specialist and sub-agents.
A 6-week real-world experiment using an open-source desktop agent shell with a three-model split (Haiku triager, Sonnet reviewer, Opus executor) reports a 64% cost reduction and details failure modes like context bloat and runaway sub-agents.
Shann Holmberg describes a structured approach to building an AI agent company within an agency, using a central brain (gBrain), an orchestrator agent (Hermes), and narrow-scoped specialist agents for different departments, with isolated client pods to prevent context leakage.
The article distinguishes between reusable knowledge (durable context) and operational memory (task state) as essential components for building proactive AI agents that can follow through on complex tasks.
An opinion piece arguing that adding more agents to a system is often a misguided fix for reliability issues, and that a single well-designed agent with better context, tools, guardrails, and evaluation is usually superior.
The author shares pitfalls from building a shared decision log for AI agent teams, including race conditions exposed by faster models, unreliable contradiction detection with cosine similarity, and challenges in testing multi-agent promises.
The author explains why they switched from single agents to chained workflows for AI tasks, citing improved reliability and easier debugging despite higher upfront complexity.
A Python learning repo that reverse-engineers Claude Code-style agent architecture through 23 incremental sessions, covering planning, subagents, context management, and more.
A Memory Curator Agent pattern separates memory governance from worker agents, improving the quality and relevance of durable memory by having a dedicated agent evaluate and route memory events into appropriate scopes.