'95% Reliability Isn't Good Enough': Where Multi-Agent AI Breaks at Scale

“95% reliability isn’t good enough” — that’s Salesforce’s framing for why order returns, compliance checks, and other high-stakes processes can’t run on probabilistic agent routing without guardrails, in a new post on where companies get stuck scaling from one AI agent to many. It names two failure patterns directly: a monolithic “jack-of-all-trades” agent that degrades as domains are added, eventually hallucinating or misrouting requests — with troubleshooting requiring engineers to comb through instruction blocks running thousands of lines — or the opposite fix, splitting into narrow specialist agents, which improves per-agent reliability but fragments the user experience and complicates cross-agent debugging. Salesforce’s proposed answer is a single orchestrator sitting as the entry point across every deployed agent: it decomposes incoming requests, identifies intent, routes each piece to the right specialist, and consolidates the responses, with named features including intelligent task routing based on agent descriptions, planned support for A2A-compliant third-party agents from Google and AWS, an “Agent Script” for declaring mission-critical routing paths explicitly, and turn-by-turn session tracing for observability.

The reliability framing is the same one Salesforce used to introduce multi-agent concepts to a much smaller audience a month earlier: its small-business explainer cited Salesforce research showing organizations already run an average of 12 AI agents, with adoption projected to climb 67% within two years, framing each additional agent as “a highly capable new hire” that acts without micromanagement. That framing works for a lean team’s first few agents; this week’s post is Salesforce acknowledging what breaks once the count climbs into the dozens — the orchestration layer that was optional at 3 agents becomes mandatory somewhere past 12. For sales orgs already running Agentforce or evaluating multi-agent deployments, the practical signal is to budget for an orchestrator before hitting the wall Salesforce is describing, not after.