Panel: Shipping Agentic Experiences - Going Beyond Pilots to Production
summary
Panel details why so many agentic projects fail and which concrete fixes scale: an open‑source agent Goose now powers millions of merchants, supports 20 LLM providers and hundreds of models, and proved useful as an MCP client; production risks include a coding agent deleting a DB and Gartner’s 40% by 2027 scrap rate; deploy patterns include GraphQL subscriptions for long 10+ minute context windows, MCP tooling over big prompts, token rotation via a token vault, and buying FGA/identity primitives rather than rebuilding them.
AI-written summary · about
chapters (10)
- 0:00Agentic AI production challenges
- 1:30Clarity in agentic projects
- 8:57Security for agentic workloads
- 12:11Scaling agentic experiences at Meta
- 17:10Speed versus correctness in agent architecture
- 24:16Build versus buy for agentic systems
- 27:28Context accuracy strategies
- 32:48Shipping agentic systems with safeguards
- 36:41Agent identity in production
- 43:19Pilot to production advice
related talks
38mPanel Discussion: AI Needs a Foundation That Never Fails
01h 23mWorkshop: Apollo Router & MCP: A Modern Agentic Development Approach
27mPanel: APIs for AIs
24mScalable GenAI in Production
30mKEYNOTE: Building Reliable Data infrastructure for the Agentic Era
33mPanel: Industry Impact - Retail
56mPanel: Scaling AI successfully
17mCrafting an Enterprise‑Scale Agent Mesh