The first agent demo at a company is usually delightful. Someone gives a model a tool, asks it to complete a task and watches it work. The hundredth agent is less serene. Two agents want to change the same infrastructure. A workflow crosses six systems with different permissions. A model produces code faster than anyone can review it. The organization discovers that the hard part was never getting one agent to do something once.
Apple’s multi-agent SRE scenario captures the problem in miniature. A remediation agent wants to add capacity during a traffic spike. A cost agent sees idle infrastructure and wants to remove it. Both are performing their assigned jobs correctly, and together they are about to make the incident worse. Resolving that collision requires shared state, priority rules, conflict detection and a clean path back to human control. Better prompts cannot decide which organizational objective should win.
Meta sits at the other end of the scale curve. Neelakshi Soni’s talk draws on distributed systems at Meta and Amazon that process millions of requests per second. At that level, AI infrastructure is constrained as much by attribution pipelines, configuration, experimentation, latency and operational complexity as it is by model capability. The agent is only the visible tip of a large and rather unromantic engineering system.
AT&T makes the organizational scale even more tangible. Its HR front door is not a chatbot pointing 130,000 employees toward help pages. It coordinates more than 100 agents, each mapped to a business process and connected to live enterprise systems through public and bespoke MCP servers. Some workflows pause for an employee and later resume. A task engine with controls designed to support SOX compliance keeps the process from dissolving into an elegant but unauditable conversation.
The same pattern appears across very different companies. Block is building a shared workplace in which people and agents carry identity, context and permissions across tools. Datadog is using agents on the critical path of code review at 10,000 pull requests a week. FINRA is turning one-off agent experiments into reusable, observable SDLC capabilities. Capital One’s session looks at applying agents to FCRA-regulated data pipelines, with an immutable audit trail built into the design.
These are not stories about companies finding a magically smarter model. They are stories about companies rebuilding the machinery around the model: orchestration, identity, state, policy, verification and recovery. Once the demo works, the agent becomes a systems problem and an organizational one. That is where the interesting engineering starts.
Continue the conversation at AGNTCon + MCPCon North America
- Multi-Agent SRE: What Happens When Your Agents Want Opposite Things with Prakshal Doshi and Aditi Mewada, Apple
- Scaling AI Infrastructure Systems at Meta Scale with Neelakshi Soni, Meta
- How AT&T Is Building an Agentic Front Door for Enterprise HR with the AT&T team
- Building the Human-Agent Workplace with Bradley Axen, Block
- Don’t Merge That with Joris Bonnefoy, Datadog
- From AI Assistants to Trusted SDLC Agents with Geetha Ramachandran, FINRA
- The Autonomous Pipeline with Gokul Prabagaren, Capital One
AGNTCon + MCPCon North America takes place Oct. 22–23 in San Jose. Register and use OUTREACH25 to save 25%.
Share
Author
