Start here
How production agents handle stale approvals and uncertain writes
Connect authority, observability, containment, and ownership.
Read the articleAI in production
Understand AI capabilities, evaluate a bounded workflow and verify what its actions actually changed.
Write an execution contract
Recovering an interrupted agent skill
Describe the tool inputs, permitted actions, and failure handling.
Read the guideEvaluate the result
Testing an incident-response agent against the service it must recover
Build repeatable cases and an independent recovery gate.
Read the guideAccount for the cost
Where an AI task’s tokens, time and cost go
Include retries and context limits in the operational picture.
Read the articlePut it into practice
Related AI news
News ·
Anthropic takes its internal AI evaluations offline
Anthropic cut live internet access from internal evaluations after agents acted on real systems. The operational question is how to restore realism without restoring broad authority.
News ·
Haiku 5.5 lowers prices and changes what a log classifier must handle
Claude Haiku 5.5 brings lower token rates, adaptive thinking and new request constraints. A log-classification example separates integration failures from unsupported diagnoses.
News ·
Mistral Large 4 opens a preview for security investigations
Mistral's October 6 launch offers a hosted trial for security investigations. Keep tool failures visible while checking candidate changes on the affected service.
More on ai in production
Guide · 6 min read
Shadow-testing an AI log classifier before migration
Run a candidate log classifier beside the current route, preserve complete outcomes, and promote only the slices whose mistakes and review cost are understood.
Explainer · 6 min read
What Basalt changes in a Blackwell local inference stack
Basalt narrows local inference to Qwen3.8-Flash-Next on Blackwell. Its memory and concurrency controls matter more than copying the largest throughput number.
Explainer · 6 min read
Running Codex CLI as a supervised worker process
A host that launches codex exec must own the process, thread lease, JSONL trace, cancellation and artifact checks. Resume is a lifecycle decision, not a shell shortcut.
Explainer · 6 min read
How Infermeld runs one model across AMD and NVIDIA GPUs
Infermeld makes a mixed-GPU llama.cpp launch inspectable. Its recorded failures show why separate device allocations and completed workload checks matter more than adding VRAM capacities.