~/blog
# notes on building software, agents, and side projects.
-
How AI agents actually work — planning, tools, memory, and the loop
A first-principles guide to LLM agents: what an agent really is, the planning loop (ReAct, plan-and-execute, tree search), tool use, short- and long-term memory, multi-agent orchestration, and how A2A + MCP slot into the picture.
-
MCP — the USB-C of LLM tools (servers, transports, and patterns that work)
A practical, diagram-heavy guide to the Model Context Protocol: what an MCP server actually is, the JSON-RPC wire format, tools vs resources vs prompts, stdio vs HTTP+SSE transports, capability negotiation, sampling, and the patterns that hold up in production.
-
A2A — building agent ecosystems that talk to each other (with OpenAPI + MCP)
A deep dive into the Agent-to-Agent (A2A) protocol: how independent AI agents discover each other, negotiate capabilities, and chain tool calls across organizational boundaries. Compares A2A vs. OpenAPI vs. MCP, the message envelope, task lifecycle, streaming, auth, and a worked example.
-
Building Vaartalaap — real-time collab rooms with Yjs, WebRTC, and zero accounts
A deep dive into Vaartalaap (वार्तालाप): a CRDT code editor, canvas whiteboard, rich-text notes, mesh WebRTC video, and chat — all in one disposable room with no signup. Architecture, the bugs that mattered, and how it ships free-tier.