Articles, one at a time.
Every piece cites primary sources and adds something you can act on: a worked example, a tested command, a documented limitation. Drafts that only summarize what already exists stay unpublished.
Beyond Naive Summarization: Context Compaction with Temporal Memory
Why long-running AI assistants bleed money on repeated history, and the memory architecture (compaction, caching, extraction) that stops it.
Read →
Why Agent While-Loops Fail in Production: Build Durable Workflows
Agent while-loops lose state on every crash, timeout, and rate limit. How durable execution with Inngest AgentKit or Temporal fixes it, with code blueprints.
Read →
RouteLLM in Production: Dynamic Cascades That Cut LLM Spend up to 85%
Why AI products bleed money on frontier models, and the three-layer architecture (pruning, caching, hybrid routing) that fixes cost without killing quality.
Read →
gpt-realtime Retires Jan 2027: One API Call Audits Your Stack
OpenAI retires nine voice model IDs on 2027-01-20. We probed the live API: no warning headers, but a shutdown_date field flags 52 of 126 models.
Read →
Adding One Tool to Your Agent Wiped the Whole Prompt Cache
Effloow Lab ran 17 OpenAI API calls. Appending, deleting, reordering or rewording a single tool zeroed the prompt cache every time. One setting avoided it.
Read →
The gpt-5 Alias Points at a Model OpenAI Deletes Dec 11
We asked the live OpenAI API which model answers when you write gpt-5. It is the exact snapshot being deleted on December 11, 2026.
Read →
Your Prompt Tooling Has a Deadline Your Monitoring Can't See
Two vendors are switching off prompt and eval tooling. We probed the live APIs, found no warning headers, and built a scanner that dates every hit.
Read →
Instrument Multi-Step Agents with OpenTelemetry GenAI Conventions: Tracing That Survives Production
A practical guide to instrumenting multi-step LLM agents with the OpenTelemetry GenAI semantic conventions, so a 2 AM failure becomes a readable trace instead of a support ticket. Includes a runnable local demo with no paid API keys.
Read →
Your Next MCP Server Could Be the Breach: A Supply-Chain Vetting and Quarantine Pipeline for Third-Party Tool Servers
A practical vetting and quarantine pipeline for third-party MCP servers: scan install scripts, verify provenance, and run candidates in a secret-free sandbox before they ever touch your agent stack.
Read →
MCP Tool Poisoning Is Now an OWASP-Listed Attack: Build a Description-Hygiene and Provenance Auditor for Your MCP Servers
MCP tool poisoning — hidden instructions in tool descriptions that hijack agents — is now an OWASP-listed attack. Build a fully local description-hygiene and provenance auditor you can hand to enterprise security reviewers.
Read →
OpenAI Spend Limits Return 429: Your Retry Logic Will Make It Worse
OpenAI hard spend limits reject requests with a 429. Lab evidence on why default client retries treat it as a temporary blip, and the one-line fix.
Read →
Your AI Gateway Blocked the Request. Did You Still Pay?
We measured what OpenRouter returns when a request is refused, and whether refused calls still cost money. Ten blocked calls moved the bill by $0.00.
Read →