Uber's MCP gateway and Salesforce's $2.7bn Cresta deal show the plumbing phase of enterprise agents

The Uber build is the more instructive of the two. InfoQ reported on 2 November that the gateway separates a registry from a proxy. The registry holds tool definitions, schemas, ownership and enablement settings. The proxy translates MCP requests into HTTP, gRPC or TChannel calls and converts the answers back. Uber's engineers describe it as a microservice layer. It exposes existing APIs as MCP tools without touching the services underneath.

That is a governance problem wearing an architecture costume.

Uber's AutoCrawler scans internal interface definition registries. It pulls API definitions out of Protocol Buffers and Thrift schemas, then uses a language model to draft tool descriptions. Newly found tools stay switched off until the owning team reviews and enables them. Uber's internal access control system enforces authorization at both server and tool level. The gateway also redacts personally identifiable information from tool responses. For third-party services including Jira and Google, a separate service swaps internal user tokens for third-party authentication tokens.

Reactions gathered by InfoQ suggest the design is familiar ground. Ajay Sharma, a principal software engineer at Microsoft CoreAI, wrote on LinkedIn that Azure API Gateways "has been doing this for a while now". He called it "a recipe for order to the chaos that LLMs come with." Kelly Goetsch, president at Pipe17, said in a LinkedIn post that Uber's gateway "looks identical to gateways we've built over the decades for APIs and later GraphQL." InfoQ also notes Uber built Omni MCP for incremental discovery. Loading thousands of tool definitions into an agent's context is inefficient.

Salesforce buys the conversation layer

Salesforce's move is the money story. The company said on Monday that it had signed a definitive agreement to buy Cresta, whose voice agents handle customer calls end to end. Salesforce did not disclose a price. It agreed to pay $2.7bn, according to a person familiar with the deal, as Valida Pau and Kevin McLaughlin reported for The Information, per TNW, which said it had not independently verified the figure. The deal should close in the first quarter of Salesforce's fiscal year 2028, subject to regulatory approvals.

Cresta is nine years old and started at Stanford University's AI Lab, The Information reported. Salesforce lists Marriott, Comcast and Alaska Airlines as enterprise users. "Voice is where service is hardest to get right, and where it matters most to customers," said Kishan Chetan, EVP and GM of Agentforce and Service Cloud at Salesforce. Cresta chief executive Ping Wu said uniting the platform with Salesforce would embed "real-time conversation intelligence and agentic experiences" into Service Cloud.

The deal follows a $5.7bn agreement for Braze on 20 October and plans to buy Rune, a national security software company, TNW reported. The pattern is consolidation of the application layer, not the plumbing.

The humans still in the loop

Two older items bracket the week. CNBC reported on 31 October that foundation models from Meta and OpenAI still cannot reliably act across the web, particularly on phone calls. Startups including Marissa Mayer's Dazzle AI, Duckbill and Wajo AI route failing tasks to human contractors. Duckbill's Meghan Joyce described a model "not dissimilar from Uber, where we boil tasks down into atomic bite-sized pieces." Mayer said models "can still lack common sense."

MIT Technology Review's custom content arm reported on 2 November that 48% of insights generated by analytics agents at surveyed organisations lead to actions that improve the business, up from 37% a year earlier. Lack of business context was cited by 55% of executives as the main reason for poor insight quality. Another 52% said their agents generate too many insights to evaluate. The same outlet reported on 3 November that researchers are pushing "reverse game theory", or mechanism design, as a way to keep agents on the rules rather than trying to inspect their internals.

Cost is the other brake. Y Combinator chief executive Garry Tan spoke on Walter Isaacson's The Builders podcast, in an episode TechCrunch reported on 2 November. He said he spends "about maybe $100[,000] to $200,000 a year on tokens". He called that spend "a crazy privilege that not everyone can do."

Elsewhere the tooling market keeps filling in. Momentic, a YC W24 company, launched on Hacker News on 2 November with an AI QA agent that writes end-to-end tests in YAML and runs them in CI. Roleo, founded by Veriff co-founder Janer Gorohhov, raised €2.9m led by Practica Capital, Tech.eu reported on 2 November. And Nvidia's blog on 3 November described city deployments of vision and service agents, including ITMAX's SamurAI V2, which it says will support 17,000+ cameras by 2027.