The AAIF community is an open and transparent community where all are welcome to participate. It is a place where Agent builders can collaborate and share

Worldwide
James Barney (MetLife) on tying token spend to cost per business outcome, and why a cheaper model you have to run twice can double the bill.
1
1
412
AAIF Community retweeted
AWS has ~16,000 APIs, so its MCP server runs in code mode. @JamesWard (@AWS) on why agent-written Python is hard to review, and the restricted, schema-validated DSL he's testing instead.
2
2
8
714
AWS has ~16,000 APIs, so its MCP server runs in code mode. @JamesWard (@AWS) on why agent-written Python is hard to review, and the restricted, schema-validated DSL he's testing instead.
1
2
394
A server that lists its skills and serves them as resources, so tools ship with instructions for using them. Ola Hungerford of @Nordstrom on Skills over MCP, dependency gaps, and the context bloat that returns at hundreds of skills.
2
2
431
WebMCP moves tool calling into the web page, scoped to the page you are on and running in your logged-in session, instead of an agent screenshotting and searching the accessibility tree. Angie Jones on why that matters and what still blocks adoption.
1
1
441
"As a company we are spending $2 per million token. What does it mean? It doesn't mean anything." Kuntal Patel and Abhinav Lad of Palo Alto Networks on why per-token cost is a weak KPI, and what they measure once agents start looping on their own.
3
1
440
Ambud Sharma of @Pinterest maps AI efficiency to five layers: hardware, capacity, inference engine, model and quantization, and routing and governance. The first two are immutable once bought, so no tuning higher up recovers a wrong order.
2
1
467
Drew Breunig on labs training their harnesses into the models: system prompts shrink each release because last cycle's hotfixes get post-trained in. Reliability up, diversity down. Anything that doesn't look like Claude Code fights the weights.
2
484
Billing alerts run a day or two behind. Agentic loops and retries can burn a budget faster than that. Brent Eubanks, FinOps architect at @Wayfair, on tracking cost per step from observability logs and alerting when one step moves 10 to 15 percent.
1
1
447
Most FinOps work right now is about slowing AI spend down. Brent Eubanks wants the opposite: "I want to enable 10X to 1000X more token use over the next two years. That's not what we're worried about." Wayfair's FinOps Architect episode drops today!
419
Jason Liu on building Codex skills: have the agent read the last 100 pull requests two teammates reviewed, extract what each one cares about, then run that as a reviewer on every new PR. Onboarding docs, not prompts.
1
2
626