Signal Brief #11: Salesforce’s Headless 360 for AI agents + 4 more

SIGNAL BRIEF

Issue #11 · Wednesday, August 26, 2026 · 5 AI tools for B2B SaaS product managers

Salesforce on Tuesday significantly expanded Headless 360, its effort to transform every Salesforce cloud from an application into reusable enterprise capabilities that any authorized AI agent can discover and use through open standards. The premise, in the company’s own framing: AI agents are rapidly becoming the new interface for enterprise software — yet most organizations still cannot let agents act safely across the business without custom integrations, duplicated business logic, and fragmented governance. The centerpiece is the Headless 360 MCP Server, now in open beta, which lets agents running in Agentforce, Claude, ChatGPT, Cursor, and other platforms dynamically discover, understand, and invoke Salesforce capabilities in real time. Because it is metadata-aware, agents don’t just find endpoints — they understand the relationships, permissions, workflows, validation rules, and governance behind them, and every action inherits the identity and permission model the customer has already built. Also shipping: the Data 360 MCP Server, now generally available, exposing nearly 200 APIs so agents can build semantic models, transform data, generate calculated insights, create audience segments, and activate campaigns in natural language; a generally available Slackbot MCP Client connecting Slackbot to Salesforce and more than 20 partner apps including Atlassian, Box, Canva, Docusign, Notion, and Zoom; a repository of more than 100 reusable Agent Skills; a generally available multi-framework for building React apps with native Salesforce authentication; a headless experience layer in open beta; and MuleSoft and Informatica capabilities exposed as governed MCP services. Distribution reaches the Anthropic, OpenAI, Google, and AWS surfaces through AgentExchange.

Why it matters for PMs: This is the clearest statement yet from an enterprise incumbent that the application UI is no longer the primary interface to its own product — the governed capability layer is. Two implications for B2B SaaS PMs. If you build on Salesforce, the objects, flows, and validation rules you maintain are now product surface that third-party agents read and invoke; metadata hygiene becomes a product-quality issue, not an admin chore. If you sell beside or against it, note the strategic bet: the platform that owns the trusted capability layer — identity, permissions, and business logic intact, reusable by any authorized agent — owns the agentic era, and Salesforce is pushing those capabilities into every major AI surface at once. The pattern worth copying wherever you build: expose capabilities with metadata-aware discovery, so agents inherit your permission model instead of forcing a new one. The bar is now explicit: capabilities that an authorized agent cannot discover and invoke increasingly don’t exist in the workflows where your customers’ work gets done.

Microsoft announced Monday that starting in September, roadmap content for Dynamics 365, Power Platform, and Dataverse moves into a single AI at Work roadmap — and with it, the company is retiring the twice-yearly release-wave model that has defined enterprise software disclosure for years. Instead of holding features for a scheduled wave, Microsoft will publish capabilities as soon as plans are committed and ready to share; each item then progresses through In Development, Rolling Out, and Launched, with status and rollout information updated as plans evolve. The roadmap is built for planning, not just browsing: views can be filtered by product and cloud environment, exported to CSV, subscribed to via RSS, and searched by feature ID. The most interesting piece for product teams is the Release Communications MCP Server, which pulls roadmap data directly into an organization’s own AI tools and workflows — so teams can generate customized roadmap views and planning materials, or wire vendor roadmap data into their own tooling, by prompt instead of portal. The transition runs September through November: content with public preview or GA dates of June 1, 2026 or later moves to the new experience, and by November 15, Release Planner retires and the AI at Work roadmap becomes the primary destination for public roadmap information across Microsoft’s business applications.

Why it matters for PMs: Release waves were a coordination device: vendors got a marketing moment, customers got a predictable planning cadence. Microsoft is conceding that continuous shipping has outgrown that cadence — features land when they’re ready, so disclosure must be continuous too. If you sell into enterprises, treat this as a preview of what your customers will start expecting from you: an always-current roadmap, not a twice-yearly PDF. Note that the bar rises past human-readable pages: with CSV, RSS, and now an MCP server, Microsoft’s roadmap is becoming data that its customers’ AI tools consume and reason over. That is worth a conversation about your own roadmap surface — is it structured, current, and machine-queryable? Internally, the In Development → Rolling Out → Launched lifecycle with continuous status updates is a clean schema to copy. One caution: continuous disclosure removes the forcing function that made organizations plan together twice a year; if you adopt the pattern, replace that alignment moment deliberately instead of letting it evaporate.

Crescendo on Monday launched its Customer Experience Platform, an AI-native system that consolidates the categories the CX software industry has sold separately for decades — contact center as a service, ticketing, workforce management, quality assurance, voice of the customer, and knowledge — and runs all of it with specialized agents. Five agents sit on top: Concierge resolves customer requests across phone, chat, email, and messaging and hands off to a human with the full thread intact; Agent Assist gives human specialists decision packets rather than raw transcripts; Applied Insights explains why the numbers moved, not just that they did; Quality scores 100 percent of conversations; and Workforce forecasts volume and staffs at five-minute granularity. Underneath, the Agentic Foundation handles optimization (finding coverage gaps in live conversations and closing them daily after go-live), integration (connecting backend systems from a natural-language prompt in minutes, so agents can issue the refund or change the order, not just answer), simulation (pressure-testing every change against a thousand simulated customer days of peak volume, bad data, and edge cases before it reaches a real customer), and knowledge (watching live conversations for what the knowledge base misses or gets wrong, and fixing it). The headline capability is recursive self-improvement with human oversight: Quality scores every conversation, Insights traces failures to a root cause, Optimization drafts a fix, Simulation validates it — and no change reaches a customer without human approval. Crescendo reports roughly 70 percent of issues resolved from day one, dissatisfaction falling from 5.8 percent to 0.68 percent in under a year, backend integrations typically in under 30 minutes, and full production in as little as 30 days. Co-founder and CPO Tod Famous frames the boundary: self-improving does not mean self-authorizing.

Why it matters for PMs: Two strategic moves in one launch. First, category consolidation: once agents can run the work, the historical boundaries between CCaaS, ticketing, WFM, and QA stop making sense — the data and the improvement loop want to live in one system. If your product is a point solution adjacent to a workflow that an AI-native platform like this can absorb, that is a positioning question worth asking now. Second, the improvement loop is the claimed moat: every resolved conversation scores the system, every failure drafts its own fix, and the gap between self-improving operations and plateauing ones compounds monthly. The governance design is the part worth copying even if you build your own version — measure 100 percent, trace to root cause, draft the fix, put simulation between the draft and a real customer, and keep a human on the trigger. If you own or influence the support experience at your company, this is the template for turning it from a cost center into a system that gets better on its own.

Auxia on Monday launched Agent Studio, an early-access product where marketers engage AI agents to plan, build, and launch campaigns across the team’s existing stack. The pitch is a control plane for marketing: describe what you want in natural language and agents plan, build, execute, and optimize campaigns across the ESPs, CDPs, and other systems you already run, once or on a schedule — campaigns launching in days, not weeks, is the claim. Three capabilities anchor it: a marketing context graph, a shared organizational graph of brand voice, past campaign performance, and business context that every agent draws on so outputs aren’t generic; end-to-end agentic workflows that handle campaign creation, analysis, and recurring operational work like weekly business reviews; and a collaborative workspace where agents collect feedback, reconcile conflicting requests, summarize proposed changes, and route approvals across legal, brand, creative, analytics, and operations. Example flows include agents detecting a brand-signal spike, generating the audience, and recommending actions for approval; rebuilding underperforming lifecycle programs from an approved creative bank; and producing localized campaigns that regional reviewers approve directly. The launch comes as Auxia crosses 200 billion cumulative autonomous personalization decisions — double its March figure — processing more than 500 million decisions daily for customers including Atlassian, Comcast, and Docomo. Agent Studio deploys standalone or beside Auxia’s Decisioning engine. Co-founder and CEO Sandeep Menon: “Lifecycle marketers aren’t blocked by talent or ideas. They’re blocked by execution.”

Why it matters for PMs: The interesting claim here is not that agents can write campaigns — it is that execution and coordination, not creativity, are the binding constraint. Menon’s diagnosis maps straight onto product orgs: half the week goes to building by hand and chasing approvals across five teams. Three patterns worth stealing for any B2B SaaS team building agent features. A shared context graph as the foundation of output quality — brand voice, past performance, and strategy loaded automatically instead of re-prompted every session. Approval routing as a first-class capability: an agent that reconciles conflicting stakeholder requests and routes changes through legal, brand, and analytics is doing organizational design, not just task automation, and that is what makes agent work safe to expand. And execution paired with decisioning: agents propose, a decision engine personalizes, humans approve. Watch the category dynamics too: Auxia is moving from an engine (decisioning) to a workspace (where the work happens) — the same surface-grab Slack, Productboard, and others are making. Whoever owns the workspace owns the workflow.

Reo.dev on Monday launched the Agent Intent Gateway, billed as the first platform designed to capture purchase intent generated by AI agents evaluating developer tools. The problem it targets: coding agents such as Claude Code, Codex, and Cursor increasingly shape how developers discover, compare, and adopt software — and those agents evaluate products by reading documentation, hitting APIs, and querying Model Context Protocol servers rather than visiting websites. Traditional intent signals — page views, doc browsing, form submissions — go dark on that journey, leaving vendors unable to see who is evaluating them, which capabilities are being explored, or whether they are making a shortlist. The Gateway acts as a proxy layer between AI agents and a company’s MCP server, surfacing which agents are evaluating the product, the documentation and capabilities they access, how frequently they return, and the depth of their evaluation — so sales, marketing, and RevOps teams can act before traditional signals appear. It extends Reo.dev’s existing developer-signal platform, which already captures activity from GitHub, package managers, documentation usage, product interactions, and technical communities. The company has raised $15.3 million from Elevation Capital, Heavybit, and India Quotient, and counts more than 200 DevTool companies among its customers, including NVIDIA, LangChain, ElevenLabs, and Temporal. Co-founder and CEO Achintya Gupta: “Companies can no longer rely only on website traffic or demo requests to understand buyer intent.”

Why it matters for PMs: Whether or not you buy this specific product, the premise is a roadmap input: a growing share of your product’s evaluation will happen inside agents that read your docs, try your API, and query your MCP server — and today that evaluation is invisible to your analytics. Three concrete implications for B2B SaaS PMs. First, documentation and API surfaces are now sales surfaces: agent-readable, accurate, complete docs are a conversion asset, not just a support cost. Second, expect “agent analytics” — who is evaluating you, what they probe, how deep they go — to become a recognizable product category the way digital analytics was a decade ago; early entrants get to define what the signals mean. Third, the funnel you measure today (visit, trial, demo) quietly shrinks as agent-mediated evaluation grows; if you aren’t instrumented for it, you will see the pipeline dip before you see the cause. The place to start is small: make your MCP server and docs excellent, and begin asking how you would know an agent is shortlisting you.