ASYNC
Clinical Guardrails, Production Pricing and a Finance Agent That Reads SpaceX Filings
Five research items from the week of July 1, 2026.
Read MoreASync by Andre Cobham
AI implementation notes. Published when there is something worth saying.
Posts cover AI strategy, infrastructure, and automation from real client work. No filler.
Latest from the Field
ASYNC
Custom Silicon, Seoul Offices and the AI Tool Your Analyst Can Actually Use
Five research items from the week of June 25, 2026.
Read More
ASYNC
Team Agents, Poisoned Servers and the Patch Window That Collapsed to Hours
Five items this week where the gap between research and production narrowed further. Enterprise agents are live, MCP has a structural flaw with real CVEs, and AI is compressing the time between vulnerability disclosure and active exploitation.
Read More
ASYNC
Open-Source Scanners, Zero-Day Windows and the PeopleSoft Breach Nobody Patched
Five items where security moved from research to active exploitation this week. Agents are both the new attack tool and the new defense layer, and real exploits are confirming the threat model.
Read More
ASYNC
A Thousand Coordinating Agents, Sparse Attention and the Fact Nobody Verified
Five research items from the past 24 hours: agents coordinating at scale, PI-Hunter for prompt injection red-teaming, MiniMax Sparse Attention, fact-checking with RL, and GPT-5.5 in enterprise pipelines.
Read More
ASYNC
The Memory Update Everyone Celebrated and the Problem It Did Not Solve
Four arXiv papers and one OpenAI product update all landed this week on the same gap. Agents without persistent memory spend 5,000 to 20,000 tokens reconstructing context each session. HORMA cuts that to 22% of baseline. PROJECTMEM blocks repeated failures. The substrate paper shows why a single memory score hides three separate failure modes. Ebby breaks down each finding.
Read More