Tag
A blog post arguing that using a frontier model only for planning and a cheaper model for execution is not cost-effective because reading—not editing—is the primary cost driver; duplicate reading offsets any savings.
This paper introduces AutoPersonas, an architecture for long-term persona agents that prevents 'self-locking' by separating divergence from evidence-governed absorption, demonstrating reduced repetition and improved identity continuity in simulated environments.
Argues that vector stores are unsuitable for storing business records (orders, balances, etc.) because semantic similarity does not guarantee factual correctness; recommends using SQL for structured data and vector stores only for unstructured information.
A discussion on how AI agents learn in three layers (model, harness, context) and why capturing user corrections as labeled examples is key to building self-learning agents that improve over time.
A detailed technical breakdown of how Hermes Agent's structured processing loop differs from basic chatbots, including prompt assembly, provider resolution, tool dispatch, and result evaluation.
A tweet argues that subagents are an antipattern in AI/software design.
An AI agent's reasoning step is redesigned to fuse multiple models in a panel-judge-synthesizer pipeline, with insights on which design choices actually improved performance.
The article argues that a single AI agent architecture is sufficient for complex tasks, echoing the 'one model is all you need' paradigm.
Explains the inverted architecture of a harnessed LLM agent, where intelligence is externalized into memory, skills, and protocols around a thin model core, with mediators governing interactions.
Most MCP servers are unnecessary; this article presents a framework for deciding when an MCP server is warranted, emphasizing the need for stable APIs and CLIs first.
Describes a two-layer small LLM architecture: a local always-on agent (Raven) on an RTX5080 and an online reasoning stack (Trinity Cortex) with three small models and a knowledge graph, arguing that small models are better than large frontier models for graph-based reasoning.
The article compares two patterns for deploying AI agents in the cloud: directly in sandboxes vs decoupling components. It explains the limitations of the sandbox approach due to cloud failures, and highlights Anthropic's Claude Managed Agent as a solution that decouples session store, agent runtime, and sandbox for resilience.
A practical guide explaining why naive multi-agent systems fail and how to build coordinated AI agent teams using Builder, Judge, and Manager roles with clear handoffs and verification.
Monday.com's Sidekick AI agent transitioned from a multi-agent architecture (200+ tools) to a deep agent architecture, solving context pollution and cost issues through three-layer tool discovery, priority delegation, code writing, and self-healing mechanisms.
This paper critiques current AI agent systems, distinguishing between agentic (external scaffolding) and agentive (internalized) systems, and proposes the Goal-Identity-Configurator (GIC) architecture for general-purpose agent models with endogenously developed capabilities, along with insights on safety and controllability.
This article reviews the design highlights and shortcomings of the OpenClaw Agent framework, and shares the author's experience in designing a better agent framework, FastClaw, emphasizing principles such as cloud-native, lightweight, and multi-tenancy.
A comprehensive practitioner's guide covering the full stack of building autonomous AI systems, from foundational transformer architecture to advanced agentic topics like multi-agent coordination and production deployment.
A tweet summarizing reflections on the article "How modern browsers work", emphasizing the value of the browser as a modern operating system, and providing 5 key insights for front-end and Agent developers, including multi-process architecture, JS engine compilation pipeline, performance optimization, etc.
Sierra's Head of Product discusses their AI agent platform for enterprise customer engagement, detailing parallel model calling, no-code Journeys layer, and payment infrastructure isolation.
A paper analyzing Claude Code reveals that its effectiveness comes from a simple AI loop surrounded by a robust infrastructure for tools, safety, memory, and recovery, rather than a complex AI brain. The study emphasizes that autonomy increases the burden on infrastructure.