Blog
Latest news and insights
Hybrid Multi-Agent Systems, Design Principles and Applications
Placing LLM-based agents uniformly across every stage of a pipeline drives up token cost and latency while eroding reproducibility. A Hybrid Multi-Agent System solves this by placing reasoning-based agents where judgment is required and deterministic functions where inputs map predictably to outputs. This post grounds the concept in how academia and industry actually define it, then walks through a spec-driven pipeline for automated web application development as a representative use case.
Reasoning and Acting, The ReAct Loop Behind AI Agents
Reasoning alone can't ground itself in current information, and an agent that only acts lacks the sophistication to plan and carry out a long-running task. ReAct ties Thought, Action, and Observation into a single loop that solves both at once, and it remains the default for single-agent design in 2026. This post traces where ReAct came from, its standard architecture, and what problems the loop actually solves.
The MCP Gateway, A Deep Dive
An AI agent running in an isolated environment needs credentials to call external tools, but the moment those credentials are placed inside the isolated environment, the isolation itself stops meaning anything. This post looks at the MCP gateway as the structural answer to that dilemma, its core mechanisms, and the protocol-level changes reshaping it in 2026.
Understanding How AI Models Put Their Parameters to Work
Split AI models by how they activate their parameters and three families fall out — Dense, which fires every parameter on every token; MoE, which fires only a slice of them; and SSM, which carries context forward as a compressed state instead. This post looks at what problem each one was built to solve, where each performs best, and how they trade off against each other.
From Query to Module — A History of LLM Cascade Routing
Cascade routing — deciding what to pick between a strong model and a cheap one, and when — started as a per-query binary choice and has since pushed that decision down into execution trajectories and pipeline modules. This post traces what RouteLLM, SWE-Router, and LLMSelector each set out to solve, how they solved it, and what's still left open.
The Software Development Harness Is Turning Into an Agent
What actually makes a coding agent work isn't the model — it's the harness wrapped around it. This post traces how the coding-agent harness grew from tool-calling and feedback loops into an infrastructure layer that now spans multi-agent orchestration and isolated execution environments.
Reading the A2A Specification — What It Covers and What It Leaves to Implementers
A2A (Agent-to-Agent) standardizes how opaque agents — ones that cannot see inside one another — delegate work and hand results back, and v1.0 brought full functional equivalence across three transports and a card signing mechanism into the specification proper. This is a pass through the whole spec, drawing a clear line between what the protocol takes responsibility for and what sits outside that boundary as the implementer's own problem to solve.
Defining the LLM Adapter — Architecture and Technical Considerations
An LLM adapter is not a thin wrapper around a provider API. It is the component that normalizes failure semantics, streaming semantics and round-trip invariants alongside the happy-path data structures. This is what actually goes wrong when you put OpenAI, Anthropic, Google and a self-hosted vLLM model behind a single layer, from the definition through the architecture to how you verify it.