Models are a component. Trust comes from the system around them.
A single agent with one big prompt hits an accuracy ceiling the moment the same input has to produce the same answer tomorrow.
Palantir FDEs @PatelVishup and Anchey Peng gave a great talk on building the system around the models to make this to work in production at scale.
This IRS workflow turns a 1,000-page rulebook into a decision tree of deterministic rules and LLM nodes.
Actions get staged for a human. Every correction becomes a new eval.
Full Talk here: https://t.co/vzeJvww4Rg
Durable, interruptible, long-running agents enable enterprise autonomy.
At DevCon 5, AIP Development Lead Matt Hawes and Forward Deployed Engineer Vishesh Patel launch Orchestrator — a new execution framework for long-running agents and automations.
@JKBOGEN We need to win. Hot take but not playing the starters this week might cost us, we win and we get the commanders, lose and we could get the lions….
@HoldenCantor We need to win. Hot take but not playing the starters this week might cost us, we win and we get the commanders, lose and we could get the lions….