Presents a GAN-inspired three-agent architecture (planner, generator, evaluator) for long-running application development, demonstrating how separating generation from structured evaluation with sprint contracts dramatically improves quality over single-agent approaches.
OpenAI's architecture for turning isolated model calls into sustained agent workflows by combining the Responses API with shell tools, persistent containers, and native context compaction.
Introduces three advanced tool use features — tool search, programmatic tool calling, and tool use examples — that enable Claude agents to dynamically discover, efficiently orchestrate, and accurately invoke tools across large tool libraries while minimizing context token consumption.
A detailed reverse-engineering of Claude Code's agent architecture, revealing how a single-threaded master loop with flat message history, structured tool interfaces, async steering, and TODO-based planning delivers controllable autonomy without multi-agent complexity.
Six battle-tested context engineering principles from Manus's production agent — including KV-cache optimization, logit masking over tool removal, filesystem-as-memory, attention manipulation through todo recitation, preserving error traces, and breaking few-shot pattern imitation.
A practical breakdown of memory architecture for multi-agent systems — covering four memory categories (working, episodic, procedural, semantic), implementation approaches from file-based context to MCP-backed databases and RAG with pgvector, plus advanced techniques like memory distillation and semantic conflict resolution.
Google's comprehensive whitepaper distinguishing agents from plain models, explaining how tools, orchestration layers, and cognitive architectures combine to create capable AI agents.
A comprehensive overview of the major design patterns for building AI agents, from single-agent architectures to multi-agent collaboration, with clear diagrams and practical trade-offs for each pattern.
Lilian Weng's seminal deep-dive into the architecture of LLM-based autonomous agents, covering planning (CoT, ToT, self-reflection), memory systems (sensory, short-term, long-term), and tool use as the three pillars of agent design.