AI Agents

A model is one component of a working system, not the system itself:

This diagram is a decomposition, not a claim that every product uses the same architecture or vocabulary.

Reading paths

Agent architecture

• Agent harness defines the control plane, permissions, state, and recovery boundary.
• Bounded loops explains progress, stopping, verification, and escalation.
• Context engineering decides what enters the active model window.
• Agent memory and retrieval explains which evidence and context to retain, and how to retrieve relevant information for the current task.
• Tool contracts defines arguments, effects, errors, retries, and authority.

Capability packaging

• Agent Skills packages an inspectable, reusable procedure.
• MCP adds a live client/server integration when interoperability justifies the extra boundary.

AI-assisted practice

• Software design for AI applies design concepts, domain language, tests, and deep modules to generated code.
• Learning with AI keeps diagnosis, recall, and next steps under learner control.
• Personal knowledge systems separates capture, private understanding, evidence, and public writing.

Evidence and current choices

• Evidence and Bias separates specification facts, vendor claims, benchmark results, observations, and preferences.
• AI Models provides dated comparisons of models, prices, runtimes, allowances, and products to narrow a choice.

Questions when combining components

One tool may serve several roles; each role does not require a separate product.

Question	Details to check
Which actions need permission?	Where authority rules are declared and what enforces them
Which steps are worth reusing?	Whether ordinary documentation is enough or a callable skill helps
How should tools connect?	Capabilities, permissions, and failure modes of local commands, MCP, or another API
What happens after a failure?	Which state can be recovered and who decides to retry or stop
Does a comparison apply to my task?	Whether product versions, test tasks, and operating conditions match

Keep the use case in view

These notes focus on terminal and coding tasks. Browser agents, embodied systems, data labeling, non-English use, organizational governance, and human-computer interaction receive less coverage.

Use provider documentation for interface definitions. To judge whether a feature is useful, safe, or better in your workflow, look for relevant task results and failure cases.

My reading · Opened notes are remembered in this browser.