On August 26, 2026, Salesforce and Anthropic jointly announced Claudeforce. Major financial media covered the announcement heavily at almost the same time, and the next day Salesforce’s stock surged 23%, marking its best single-day gain since 2020. Four months earlier, the same company had made another announcement at the TDX conference on April 15, 2026, unveiling Headless 360 to fully open up the entire platform to AI agents. The official statement at the time was: “Everything on Salesforce is now an API, MCP tool, or CLI command, and agents can use all of it.” Yet on that day, mainstream financial media barely gave it any coverage.
These two launches were not independent. On the day Claudeforce was
announced, Marc Benioff explicitly noted that Claudeforce runs directly
on top of Headless 360, stating: “running directly on top of Salesforce
via our new AIforce UI harness, Headless 360, Data 360, Tableau, and
Slack.” In other words, Salesforce shipped the exact same underlying
capabilities twice within four months. What launched in April was an
exposed connection pipe, while what launched in August was a fully
packaged application product. The connection pipe released in April saw
little use, whereas the product packaged in August drew widespread
attention.
Following this contrast further, the question becomes concrete. Right now, nearly every software company is trying to answer the same question: their software must be operated not only by humans, but also invoked by various agents. In form, almost everyone has converged on MCP servers. But the four months showcased by Salesforce resemble a well-controlled natural experiment. The interface was put in place early on, but interfaces alone make nothing happen. The critical conditions that actually drive adoption are hidden in the layers above the interface.
Putting the April launch back into its context at the time: the core component of Headless 360 then was Salesforce Hosted MCP Servers, which entered general availability (GA) on April 15. As an open-source communication protocol released by Anthropic in late 2024, MCP had already become the de facto standard for agents connecting to external systems, with GitHub, Slack, and Notion having already launched official MCP servers. The step Salesforce took was supposed to expose the CRM’s internal data assets, business workflow logic, and access control entirely through MCP interfaces, theoretically allowing any agent client to connect directly.
In terms of industry buzz, the April launch could not be considered a failure. VentureBeat published an in-depth analysis at the time, and Salesforce’s own developer ecosystem media followed up closely. But looking at actual adoption, over the subsequent four months, the only meaningful traces found across the internet were essentially tutorials—roughly 8 to 12 third-party walkthrough articles and demo videos repeatedly explaining how administrators should configure parameters and how users could connect into Claude. Yet across public channels, not a single enterprise could be found that actually put this official pipe into production to run real business operations.
Looking through developer discussions on Reddit, real feedback from frontline engineers fell clearly into two camps. The first camp tried to take the official path but encountered numerous obstacles in practice. A popular thread listed ten specific issues, the most typical being that even though the OAuth flow showed successful authorization, the system still threw errors denying access; after extensive troubleshooting, it turned out that the server URL in the official documentation was missing a path segment. The second camp of developers bypassed the official hosted MCP entirely, writing their own gateway in Python to wrap legacy underlying APIs, citing that the official hosted MCP was “totally useless” at the time. Independent research from the capital markets also corroborated this quiet state: five days before the Claudeforce launch, TD Cowen conducted a survey of ecosystem partners, and not a single respondent cited Agentforce as a driver of recent deal flow.
The most compelling evidence actually came from the retrospectives of Salesforce’s own management. As reported by Apex Hours, Patrick Stokes, President of Applications & Marketing at Salesforce, later admitted that after TDX, many developers did start building agent interfaces against the MCP server, but then “immediately hit a wall: managing authentication, granting execution permissions to over a hundred users… every team was solving the same plumbing problem poorly, one user at a time.”
Looking at these points of friction together, the obstacles to adoption fell into two specific areas, largely unrelated to the protocol itself. The first hurdle was administrative and operational. To get the official pipe running properly, enterprise admins had to manually complete a seven-step setup: from activating the server, creating an external app, precisely configuring callback URLs, and setting OAuth policies, to manually copying generated keys into the Claude client. Furthermore, authentication tokens were tied to each logged-in user’s individual account—for a team of one hundred employees, the admin had to repeat credential distribution one hundred times. The second hurdle lay at the business scenario layer. A bare pipe only provided agents with a collection of atomic APIs; business logic—such as how to prepare for a daily sales standup or how to assess deal health—was absent from the underlying interfaces, requiring agents to piece everything together on their own. While tutorials and discussions abounded, actual enterprise adoption remained virtually nonexistent; this gap defined the real picture between April and July.
By August, the product was officially named Salesforce in Claude, re-emerging as a plugin pre-loaded with 37 sales-specific skills. Against the friction list from April, the additions in this launch spanned exactly three layers.
The first layer is packaged administration. The core promise on the product page was: “An admin connects Salesforce in Claude once, and it works for the whole team from day one.” Enterprise permissions management no longer required configuring one user at a time; employees connected directly with their native Salesforce roles and permissions, which were inherited automatically. The company even turned this into a headline selling point: “There’s nothing new to stand up, nothing to re-audit.”
The second layer is the deep encapsulation of business semantics. The 37 built-in skills are entirely task-oriented, with specific names like daily-briefing, pipeline-review, stakeholder-map, and close-plan. Each skill corresponds to a real frontline sales workflow, embedding business knowledge—such as which fields carry the most weight and how deal health is calculated—directly into execution logic. This semantic layer differs fundamentally from underlying MCP: MCP hands an agent a feature inventory, whereas skills deliver a methodology for getting work done. There is a telling detail here: in the promotional diagrams on the official product page, the 15 publicly displayed skills are all arranged around an icon representing a skill.md file. According to the underlying specifications of the Claude plugin system, skills are essentially specific directories containing SKILL.md files—a standard that corresponds to the open Agent Skills specification that Anthropic introduced to the industry in 2025, which OpenAI subsequently quietly adopted in its own products. In other words, a leading SaaS company is directly using an open file format established by model providers to encapsulate its business tasks. It should be noted that Salesforce has not made an explicit public statement regarding this implementation detail, and the 37 file contents themselves have not been made public; however, based on plugin runtime mechanics and supporting collateral on the product page, the technical direction is unmistakable.
The third layer is tight binding with end-user distribution. If administration and semantics solve whether capabilities can be used effectively, distribution solves where users actually interact with them. This plugin is embedded directly inside the Claude interface, where sales reps already spend their daily work and communication. A summary line on the official promo page puts it clearly: “A plugin you install instead of a stack you assemble by hand.”
That said, while recognizing these additions, it is also necessary to objectively note the unresolved issues in this launch. On announcement day, the product was opened only to a limited group of pilot customers; a true public beta will not arrive until September. Pricing plans have not been disclosed, and enterprise buyers must still sign two separate contracts with Salesforce and Anthropic. As Stokes noted in an interview: “You can’t buy this on one piece of paper at the moment.” Regarding customer endorsements, Salesforce did not name a single external enterprise customer; the sole testimonial came from an internal Salesforce sales executive. There is also the matter of the stock price: August 26 coincided with Salesforce’s earnings release, and the next day’s 23% single-day stock jump was primarily driven by financial results exceeding market expectations and short covering across SaaS stocks, with earnings per share also including a $2.61 billion gain recognized from its equity investment in Anthropic. Claudeforce served more as a narrative amplifier, propelling Benioff and Dario Amodei together into CNBC’s live broadcast studio. Yet on Hacker News, where developers congregate, the discussion thread on Claudeforce finished with just 9 points. The August launch captured narrative attention, but the actual depth of software adoption in real business workflows must still wait for data following the September public beta. The two launches each represented a different angle, and the contrast between them illustrates the core issue precisely.
Looking beyond Salesforce as a single company, the entire software industry has converged on agent form factors much faster than expected. On the BG2 podcast in late 2024, Satya Nadella remarked that the very notion of business applications would likely collapse in the agent era: “I think the notion that business applications exist, that’s probably where they’ll all collapse, right in the agent era.” In October 2025, Vercel founder Guillermo Rauch put it even more bluntly: the job of tooling developers is “no longer just serving developers, but serving their agents as well,” where yesterday’s starter template was an interactive web page and today’s starter template is an agent that can take action. On the underlying protocol front, OpenAI adopted MCP in March 2025, Google followed with support in April, and Anthropic formally donated MCP to the Linux Foundation’s newly formed Agentic AI Foundation in December. By August 2026, Google’s originally spearheaded A2A protocol was also merged into the foundation. Alongside this wave of standardization, the interface layer rapidly commoditized: public registries now catalog nearly 10,000 MCP servers, with The New Stack observing that “seemingly every SaaS service exposing an MCP server now.” When 10,000 generic water supply systems exist on the market, merely connecting a pipe no longer constitutes a competitive moat.
In system architecture, the emerging industry consensus is a composite stack of multi-layered capabilities, rather than an either-or choice. Salesforce’s official plugin technical documentation outlines a clear three-tier resolution order: “skills first, then the Salesforce CLI, then Salesforce MCP.” Each layer has distinct responsibilities: skills organize business methodologies and workflows, the CLI provides deterministic local command execution, and MCP handles secure underlying connectivity and data retrieval. With each tier serving its specific purpose, no vendor chooses between interfaces and skills as an either-or trade-off.
At the same time, cautious and critical voices have persisted within the industry. LangChain founder Harrison Chase stated publicly that the current MCP-1 looks a lot like USB-1 back in the day, and that better-designed new protocols will emerge in the future. Data from the security domain provides further evidence: GitGuardian scanned and detected over 20,000 leaked secrets directly tied to MCP configuration files in public GitHub repositories; individual users have also reported mounting 58 servers and pooling over 680 tools in a single development environment. At MCP Dev Summit 2026, the industry’s own diagnosis was that MCP has entered a load-bearing phase, with open challenges concentrated entirely at the operational layer: sessions, sprawl, identity, and auditing. Identity, sessions, and auditing happen to be precisely what the protocol bypassed in its early days to stay lightweight; adding them back piece by piece now makes the protocol thicker. These criticisms do not diminish MCP’s industry standing, but they point to the same reality: the communication protocol itself is rapidly converging, and all remaining problems lie in the layers above it.
The framework left by this experiment is reusable. When seeing any
SaaS vendor announce agent support in the future, examining them across
these four levels reveals that most announcements fall short right at
the first level. 
Level 1 is the interface: does the vendor provide native MCP or third-party wrappers, and what is the underlying read/write granularity? The interface layer is a prerequisite; today, when everyone has an interface, this layer alone is insufficient to drive actual adoption. Level 2 is semantics: has the vendor distilled business rules into task-level execution units—such as structured skills, automated workflows, or standard playbooks—or does it merely offer raw APIs? Level 3 is governance: can agent actions inherit the enterprise’s existing permission hierarchy and audit logs? Can administrators complete configuration once globally, or must they authorize every employee individually? Is the confirmation mechanism for write operations mandatory blocking or adjustable on demand? Level 4 is distribution: in which high-frequency client interface are these packaged capabilities ultimately embedded, and what kind of tie-in has the vendor established with the model ecosystem? Claudeforce established its narrative in August because it filled in Levels 2, 3, and 4; the April launch saw little adoption precisely because it delivered only Level 1.
For Claudeforce itself, its actual trajectory will depend on three checkpoints arriving in September. First, with the kickoff of the public beta, whether the detailed list of these 37 built-in skills and their read/write permission granularity will be fully disclosed. Second, how the system’s default confirmation logic will be configured for write operations on critical business data. Third, when the market will see the first publicly named external enterprise customer. These three frontline feedback signals are more honest than any keynote speech.
This four-month natural experiment leaves a clear heuristic: interfaces for agents to invoke are no longer scarce; what holds real commercial value is the engineering capability to package interfaces into complete products. The encapsulation of business semantics, inheritance of organizational governance, and embedding into distribution channels—these three layers represent the minimum viable adoption unit for enterprise software in the agent era. Connecting the pipes is only the first step; only when water is packaged into containers will buyers be willing to pay.