Browse articles

Something useful with your coffee

Practical AI engineering lessons, distilled from the good stuff and written for humans who build things.

Warm editorial workbench showing code diffs, trace notes, tests, and a user correction feeding into a consolidation machine that outputs memory, test, and skill cards while noisy transcript scraps are discarded.

Agent Workflows

How To Make An Agent Learn From One Session

A practical guide to turning one completed agent session into scoped memory, tests, hooks, skills, or evals without pretending the model trained itself.

Warm illustrated engineering workbench where a tangled red failure trace flows through inspection tools, lab notes, and test cards, then returns as a clean green evaluation loop beside a secondary score gauge.

Evals

Evals Are Improvement Loops, Not Scoreboards

Learn how to turn traces, failures, datasets, and experiments into a practical eval loop for improving AI applications and agents.

Warm desk scene showing a Windows desktop application with UI Automation handles, evidence panels, and test verdict overlays for background UAT automation.

AI Agents

Headless UAT Agents Need Handles, Not Screenshots

AI agents can help automate UAT for Windows desktop apps in the background, but they need isolated workers, UI Automation handles, deterministic checks, and reviewable evidence.

Warm isometric illustration of a long-term memory workbench where cards pass through intake, scoped storage, retrieval routes, and maintenance drawers.

Agent Memory

Long-Term Agent Memory Is A Retrieval Workload

Learn how to design long-term agent memory around admission, scoped records, retrieval paths, maintenance, and forgetting instead of dumping everything into vectors.

A miniature code-review cart moves through checkpoint stations on a desk while loose instruction papers pile up beside an optional specialist review lane.

Agent Workflows

Instructions Are Not Control

Why agents skip parts of your skills, and how to make expectations stick with staged workflows, structured evidence, hooks, validators, evals, and selective subagents.

A warm futuristic agent test range shows a glowing agent core inside a transparent harness, surrounded by mission intake cards, context panels, tool gates, state checkpoints, verification sensors, and a trace recorder.

AI Agents

Your Agent Needs A Harness, Not Just A Prompt

Agent harnesses are the control systems around AI agents. Learn what they own, how the core parts work, and how to evaluate one before trusting it with real work.

A developer desk shows a neat working-state board with goal, constraints, decision, evidence, next step, and checkpoint cards while a large chat transcript and tool-output stack compress into a compact summary bundle.

Agent Memory

Short-Term Memory: Working State Without Context Bloat

Learn how to keep an agent oriented during a task by carrying useful working state instead of dragging the whole transcript forward.

Warm stylized illustration of a compact agent console loading a small launch checklist while a messy archive stays behind a retrieval gate.

Agent Memory

Current Memory: The Agent's Loaded Brain

Learn what belongs in an agent's always-loaded context, what should move elsewhere, and why memory files are guidance rather than enforcement.

A warm illustrated API contract turnstile routes structured context cards through schema, permission, guardrail, and trace checks into narrow specialist lanes, with a crossed-out org chart in the background.

AI Agents

Delegate Agents Like APIs, Not Coworkers

Handoffs, subagents, routers, and skills are useful only when the boundary is designed. Treat agent delegation like an API with scoped context, checks, traces, and ownership.

A warm illustrated developer console with a large pressure gauge, code editor, test indicators, trace charts, and budget controls, suggesting thinking effort as an engineering tradeoff.

Reasoning Models

Choose Thinking Effort Like An Engineering Budget

Low, medium, and high thinking settings are not magic intelligence levels. They are budget choices for how much internal reasoning a model should spend before answering.

Agent Memory

How Agent Memory Actually Works

Agent memory is not one database or a longer prompt. Learn how to decide what an agent should load, carry, recall, update, and forget.

Warm Coffee With Humans style illustration showing messy stacks of documents, logs, and tool cards being curated into a clean AI workbench with selected context cards, a test result, trace panel, and BrewBot.

Context Engineering

Context Is The Model's Working Memory

Context engineering is not prompt polish. It is the discipline of giving AI systems the right working set at the right moment, then keeping that context fresh, small, and verifiable.

Warm Coffee With Humans style illustration showing a tiny prompt card on a paper trimmer beside a larger transparent agent workflow budget machine with cache, tool, routing, retry, output, batch queue, and BrewBot elements.

Token Usage

Stop Treating Token Usage Like A Prompt Problem

Token costs are not just a pricing problem. Learn how to reduce AI agent spend by measuring full runs, shrinking repeated context, caching stable prefixes, batching offline work, and routing tasks by value.

Warm editorial illustration of an MCP server gate marked as a trust boundary, checking tokens, tool schemas, and context packets before allowing tool calls to reach repo, database, and deployment systems.

MCP Security

Treat MCP Servers Like Trust Boundaries

MCP makes agents useful by connecting them to real systems. That also makes every MCP server a security boundary worth designing, reviewing, and monitoring.