Agent

A system that uses a language model to plan, execute tool calls and iterate towards a goal, rather than producing a single response.

An agent holds state across steps, can invoke external tools, and decides its own control flow. This makes it capable of completing multi-step tasks, but also makes the blast radius of a compromise much larger than a single chat response. The security posture is determined by what tools are available and what permissions they carry.

Explained in depth

How agentic AI works

Follow a document assistant from a user goal through model decisions, authorised tools, observations and verification, with a visual agent loop.

Agent planning

Why interleaving reasoning with action outperforms planning everything up front, what ReAct and Reflexion actually showed, and where a plan should be allowed to change.

Agent memory

Separate context, checkpoints, execution archives and long-term memory; design admission, retrieval, correction and deletion with source provenance.

Diagnosing a failed agent run

Separate model proposals, tool attempts and real outcomes; follow an uncertain write through a redacted trace and turn the failure into a regression test.

A2A - the Agent2Agent protocol

How the A2A protocol v1.0 standardises communication between independent agents - the Agent Card for discovery, the Task as the unit of work, streaming and push delivery, and where the boundary with MCP actually sits.

Agent harnesses - the layer that owns the loop

The difference between an agent harness, an agent framework, and a protocol - what a harness actually provides (loop, permissions, sandbox, context, recovery) and why the layer names get confused.

Agent Skills - packaged instructions with progressive disclosure

What an Agent Skill actually is - a folder with a SKILL.md file, a three-level progressive-disclosure model that controls what enters context and when, and why the description field is the whole trigger mechanism.

Agent sandboxes - where unvetted code runs

The infrastructure layer that runs code an agent wrote without a human reviewing it - the isolation primitives (container, gVisor, microVM), the lifecycle a sandbox exposes, and how E2B, Modal and Daytona differ on the dimensions that actually break a deployment.