AI & agent systems5 min read
From chat to execution: building agents that finish the work
A useful agent needs more than a good answer. It needs a plan, tools, a place to work, and a way to know what happened next.
Writing & ideas · 24 articles
Notes from the workbench
Field notes on building AI agents that get useful work done, the products around them, and the developer tools I make along the way.
Developer tools4 min read
Turning repeated development work into reusable workflows
Notes from building ISTMX Skills: making practical software workflows reusable across coding agents without tying them to one stack.
AI & agent systems4 min read
Why the computer is part of the agent
A browser, terminal, files, and applications give an agent a place to act. Treating that environment as its own layer makes the system easier to extend.
AI & agent systems4 min read
Useful agent memory does not mean loading everything
Keep task state structured, retrieve relevant context when needed, and let an agent find the right memory without stuffing every document into every request.
AI & agent systems4 min read
One coordinator, many specialists
A multi-agent system works best when specialist roles serve one clear goal and a main agent stays accountable for bringing the result together.
Product notes3 min read
Research that fits into your existing workflow
Noiseless explores a simple idea: monitor a topic on the schedule you choose, prepare a useful digest, and deliver it where you already work.
Product notes3 min read
Keeping an AI code review available when providers fail
CodeCat uses a multi-provider fallback approach so a temporary issue with one AI provider does not have to stop a code review.
Building products4 min read
Building an AI product across web, backend, and mobile
An AI feature lives inside a whole product. Web and mobile interfaces, backend services, and persistent state all shape how people use it.
Backend engineering3 min read
Designing FastAPI services for AI products
A Python API around a model needs clear boundaries, validation, and predictable failure behavior before it needs more endpoints.
Python3 min read
Python type hints that help API teams
Types are most useful when they describe boundaries clearly and catch mismatches before a request reaches production.
AI systems3 min read
When an AI task belongs in a background job
Long-running agent work needs a lifecycle that survives a browser tab, a request timeout, and a temporary provider failure.
AI engineering3 min read
RAG is a data pipeline before it is a prompt
Retrieval quality depends on what gets indexed, how it is chunked, and whether search finds evidence relevant to a question.
AI evaluation3 min read
Evaluate agents by task outcomes, not confident wording
Fluent text is not proof that an agent completed the work. Evaluation should check the artifacts and constraints that matter.
Agent design3 min read
Tool schemas are product interfaces for agents
An agent can only choose tools well when names, descriptions, and input requirements make actions understandable.
AI safety3 min read
Prompt injection is a boundary problem
Instructions found in a webpage or document should not gain the authority of the user or system processing it.
Product design3 min read
Design AI error states that help people recover
Explain what happened and what someone can do next without pretending to know more than the system does.
Mobile development3 min read
Design mobile AI experiences around long-running work
A mobile app should let people start, follow, and revisit work without requiring them to keep a screen open.
Mobile development3 min read
Connecting an Expo app to a web backend
Shared contracts help web and mobile clients agree on data, while platform-specific UI keeps each experience natural.
Backend engineering3 min read
Authentication and authorization solve different problems
Knowing who made a request does not tell you whether they may read or change a particular resource.
Backend engineering3 min read
Model agent tasks so they can be recovered
A task record should preserve enough state to show progress, retry safely, and return finished artifacts after a worker exits.
Full-stack development3 min read
Keep TypeScript API contracts clear across the stack
Shared types catch client-server mismatches, but they do not replace runtime validation or stable API design.
AI product design3 min read
Stream AI responses without faking progress
Streaming can make waiting feel responsive, but the interface should distinguish generated text from completed work.
AI systems3 min read
Observability for agent workflows starts with a trace
A useful trace links the request, model calls, tools, and final outcome without exposing secrets.
Frontend development3 min read
Build accessible interfaces for AI features
Changing AI output makes status announcements, keyboard access, and reduced-motion behavior especially important.