Skip to content

Pillar I

Agentic infrastructure, made operational

The how behind intelligent systems. Not prompting tricks or model comparisons, but the practical layer that makes agents reliable enough to trust with real work: protocols, tool design, context engineering, memory systems, and the orchestration patterns that hold them together.

This pillar carries the method. The other two pillars give the work its direction and its edge. This one explains how the systems are actually built.

Implementation layers

The stack beneath useful agents, from protocol to orchestration.

Protocols

Tool calling & MCP

Tool calling is the loop that makes agents useful: provide tool definitions → model emits tool calls → application executes → tool output returns → model continues reasoning. MCP (Model Context Protocol) extends this to remote servers via JSON-RPC 2.0, giving models access to tools, resources, and prompts from any conforming server. The protocol isn't the interesting part — what you build on top of it is.

Ergonomics

Tool design as UI for agents

Tools are the agent's interface to the world. Bad tool design degrades reasoning: vague descriptions cause misuse, noisy outputs waste context, overlapping tools create ambiguity. Good tools have a single clear purpose, a precise name that implies what it does, and return only the signal the model needs for its next step. This is interface design — it just happens to be for agents, not users.

Context

Context engineering

What enters the context window at each agent step determines reasoning quality. Context engineering is the practice of shaping those inputs deliberately: write durable facts to external memory, select only relevant history via retrieval, compress verbose outputs before they enter context, and isolate tool outputs that shouldn't contaminate the main reasoning trace. Most agent failures are context failures.

Memory

Agent memory systems

Memory is what separates a stateless chatbot from an agent that accumulates understanding. The layers: in-context (fast, ephemeral), external storage (persistent, retrievable), and semantic memory (structured, queryable). The hard problem isn't storage — it's deciding what's worth remembering, how to retrieve it when relevant, and how to let old memories expire without losing what matters.

Orchestration

Control-flow patterns

Agents compose into systems through a small set of patterns: routing (classify → dispatch), parallelisation (fan out → aggregate), orchestrator-worker (plan → delegate → coordinate), evaluator-optimiser (generate → evaluate → iterate), and the foundational ReAct loop (reason → act → observe). These patterns aren't magic — they're the grammar of reliable agent behavior.

From the build log

Field guides, pattern analyses, and writing from inside the systems work rather than around it.

Read the Work