A resumable agent needs safe side effects
Saving progress helps an interrupted workflow continue, but external actions need their own identity and confirmation.
An independent technology journal
Technology moves fast.
Find the context.
Clear perspectives on AI, open software and the decisions behind them.
In focus
Choose the evidence path before the context windowFollow the evidence behind a technical choice.
A place to start.
Three reading paths01
AI agents: from permission to proof
A short reading path for deciding what an agent may do, and how to check what it actually did.
Follow the thread 02AI infrastructure: count both budgets
Connect model choices to hardware, operating cost and the work needed to run a reliable service.
Follow the thread 03Published work needs an evidence trail
Follow a source from collection to an approved draft and a page that can be checked and shared.
Follow the threadThe reading room
Worth a closer look.
An agent evaluation should inspect the work
An answer can look right while the underlying tool calls, permissions or evidence are wrong. Design tests around the whole task.
Published work needs an evidence trail
Follow a source from collection to an approved draft and a page that can be checked and shared.
AI infrastructure: count both budgets
Connect model choices to hardware, operating cost and the work needed to run a reliable service.
AI agents: from permission to proof
A short reading path for deciding what an agent may do, and how to check what it actually did.
A shared page needs its own identity
Treat the title, canonical address and preview image as part of the published article, then inspect the actual public response.
A launch demo needs a usable finish line
A public URL, a bounded task and a clear result give people a way to inspect a product after the launch post disappears from view.
A launch page should answer the next question
A useful first screen puts a meaningful action beside enough evidence to understand it. Keep the deeper explanation one step away.
A JSON schema defines the handoff
Structured generation helps software accept an answer. Domain checks must still establish that its values are supported and its proposed actions are allowed.
RSS discovery needs a source ledger
We parsed a real 18-item RSS snapshot and replayed duplicate, revised and malformed inputs to show what a source ledger must preserve.
Prompt caching starts with stable inputs
Before forecasting savings, identify the prefix that really repeats, the work a cache can skip and the boundaries that keep reuse safe.
Your evaluation set is a product specification
A useful AI evaluation explains who the product serves, which errors matter and what evidence makes an answer acceptable.
An agent permission should describe an action
Turn broad access requests into bounded capabilities: a resource, an operation, an intended recipient and an observable result.
Choose the evidence path before the context window
RAG and long context place different demands on evidence selection. Compare the questions, missing sources and update rules before choosing a pipeline.
Local AI has a hardware budget and an attention budget
Moving a model onto your own machine changes the bill and the operating work. Count both before choosing a stack.
A review belongs to an exact draft
An approval loses meaning when the underlying facts or text change. Bind the review to the version you intend to ship.
Read a release note before you upgrade
A version number is a useful starting point. A good upgrade decision needs a closer look at scope, dates and rollback.
A small server changes the architecture
Three publications do not need three always-on application stacks. Start with the resources you can actually measure.
No matching notes.
Try another topic or clear the filters.