The Agentic Enterprise AK · Morning Edition · 8 min read | Tuesday, September 15, 2026 The Agentic Enterprise Takes the Stage Dreamforce 2026 opens today in San Francisco with Salesforce putting Agentforce 360 at the center of its biggest event, a Claude keynote pairing Marc Benioff with Anthropic's Dario Amodei, and Sam Altman on the bill. The three companies shaping enterprise AI are converging on a shared stack, adoption is climbing fast, and the open question for leaders has shifted from whether to move to where to start. That makes this the most interesting week in AI this year. For one week each year the enterprise-software industry gathers in San Francisco to agree on where the future is heading, and this year the destination is unusually clear. The agentic enterprise, the idea that autonomous agents rather than static apps become how work gets done, is moving from concept to architecture. Marc Benioff has built the whole show around it, and the guest list tells you the direction is real: Anthropic's Dario Amodei shares the keynote stage to put Claude inside Salesforce, and OpenAI's Sam Altman is on the bill. The three companies most responsible for enterprise AI are lining up behind one stack. That convergence is what makes this week worth your attention. The useful posture is neither hype nor suspicion. It is curiosity with a notebook: watch where the platform is heading, note what is real today versus what is coming, and decide where your first or next agent earns its keep. | The Big StoryEnterprise Deployment |
The agentic enterprise takes center stage in San Francisco | D | reamforce 2026 opens today at Moscone Center, and Salesforce has built the entire week around a single idea: the agentic enterprise. Marc Benioff will use the keynote to position Agentforce 360, the company's unified stack of platform, Data 360, Customer 360 apps, and Slack, as the place where autonomous agents live and work. Salesforce says the show runs more than 1,600 breakout sessions and 50-plus keynotes, most of them pointed at agents. Last week the company shipped seven named agents with job titles. This week it makes the case for the operating system they run on. A year ago the agentic enterprise was a phrase in a keynote. Today it is a product line with a roadmap. |
The guest list tells you where the industry is heading. Benioff is scheduled to share the stage with Anthropic's Dario Amodei for a keynote on Claude inside Salesforce, with a Salesforce-in-Claude integration reported to be moving from limited pilot toward open beta. OpenAI's Sam Altman is also on the bill. The three companies doing the most to define enterprise AI are converging on a common architecture, and that convergence is the real headline. When the platform, the frontier model, and the runtime start speaking the same language, the integration tax that has slowed most enterprise AI programs comes down, and building an agent that touches real systems gets meaningfully easier. What is worth tracking is the distance between the demo and the deployment. Coverage going into the show notes that customer-reported ROI and production run-cost figures are still thin, which is normal for a category this young and a fair reason to stay hands-on. The good news is that the questions to ask are now clear and answerable: what a task costs, how the agent is governed, and where the first reliable win sits. This is a lean-in week, not a wait-and-see one. |
The Spearhead Take This is a lean-in week, not a wait-and-see one. Agents are becoming the interface to enterprise software, and Salesforce, Anthropic, and OpenAI are converging on a shared stack, which makes now a good time to run a real pilot and build the muscle. Watch the keynote for where the architecture is going. Then pick one workflow with a clear payback, put sensible guardrails around it, and learn by shipping. The teams that come out of this cycle ahead will be the ones who started, not the ones who waited for a perfect case study. |
| The Obvious & The OverlookedWhat the keynote will make loud, and what is worth a closer look. The Obvious Dreamforce is all-in on agents. Salesforce built the entire show around the agentic enterprise, with Agentforce 360 as the centerpiece. Salesforce The frontier labs are onstage together. Amodei and Altman are both on the bill, a sign of how tightly models and enterprise platforms are now linked. Salesforce Ben Salesforce and Anthropic are getting closer. A Salesforce-in-Claude integration is moving toward open beta as the two firms co-market the stack. Salesforce Ben | The Overlooked The stack is quietly consolidating. Salesforce, Anthropic, and OpenAI converging on one architecture is the signal that matters most for how you build. Salesforce Ben The integration tax is coming down. A shared platform, model, and runtime means less custom plumbing between an idea and a working agent. Salesforce The proof is arriving, workflow by workflow. Customer ROI figures are still early, so the advantage goes to teams that pilot now and measure their own. MarketScale Adoption is already the norm, not the experiment. A large majority of new enterprise apps now ship with an agent inside, which changes the default you are building against. Digital Applied |
| Moving PiecesFive developments worth a CIO's attention. Policy / DealsWashington bought ChatGPT for the whole government The GSA and OpenAI struck a new OneGov agreement, effective October 1, that drops the ChatGPT license fee to zero, cuts token-based usage costs 50 percent, and runs 27 months with no minimum commitment, covering the federal executive, legislative, and judicial branches plus state, local, and tribal governments. It replaces last year's symbolic one-dollar deal with something durable. For the public sector this is a genuine on-ramp: two years of stable, discounted access is exactly what a large, cautious buyer needs to move from pilots to real programs. The prudent counterpart to a deal this good is simple portfolio hygiene, keeping your data and workflows portable so today's bargain stays a choice tomorrow. BuildersAnthropic resets Claude Code limits as the market matures Anthropic ended a temporary 50 percent weekly boost on Claude Code that had run since May, replacing it with a permanent 25 percent increase over pre-May levels. Because the summer allowance was higher, the net for Pro, Max, Team, and seat-based Enterprise plans lands below the peak, and the company reposted a clearer note after developers asked for the plain math. Read it as a young market settling into sustainable pricing rather than a step back. The takeaway for buyers is a familiar one from every input you depend on: track your consumption, right-size the plan to the work, and treat model capacity as a line item you manage, not a surprise you absorb. InfrastructureA $205M bet gives buyers more choice at the wiring layer Cornelis Networks, the 2020 Intel spinoff, raised $205 million led by IAG Capital Partners and unveiled Active Compute Fabric, an open, GPU-agnostic networking layer that works alongside Nvidia's ecosystem. It pairs data transport with in-network acceleration, claims up to 50 percent less network traffic in large clusters in pre-production simulations, and arrives with a Qualcomm collaboration for rack-scale inference. Its CN5000 switch ships now, the 800-gigabit CN6000 in the fourth quarter. The enterprise upside is optionality: a healthier field of interconnect options means better performance-per-dollar and more leverage in your next infrastructure conversation. Competition at the wiring layer is how the whole stack gets cheaper. CompetitionGoogle gives its own engineers Claude, and that is a healthy sign Google has quietly opened Claude Opus 5 to every engineer through its internal Antigravity IDE, per Business Insider, easing a longstanding preference for Gemini-only internal tooling. Engineers select Opus 5 inside Google's own environment with per-user quotas, and Google notes Gemini remains its foundational internal model. The signal is pragmatic and encouraging: even the company with the strongest reason to standardize on its own model is letting teams reach for the best tool for a given job. For enterprises wrestling with single-vendor mandates, this is a useful example of a mature posture, standardize where it helps, and stay open where capability makes the difference. DataAWS and Stardog make it easier to ground agents in your own data AWS and Stardog launched a semantic layer for agentic AI that lets agents query Aurora and Redshift without custom ETL, running on Bedrock AgentCore while preserving relational context. It is the unglamorous plumbing that decides whether an agent is useful: an agent is only as good as its access to trustworthy, well-modeled enterprise data, and most early disappointments trace back to that gap rather than the model. Tools that connect agents to governed data with less bespoke engineering are exactly what move pilots into production. This is the kind of enabling layer to watch as you plan your own deployments. | On the RadarNine signals, sharpened. | Compute | Anthropic's compute commitments reached $517B. The lab has agreed to roughly $517 billion covering 14.8 gigawatts through August, a signal of how much capacity it expects enterprise demand to need. 247 Wall St | | Governance | Microsoft paired an endorsement of careful pacing with a rulebook. Satya Nadella backed deliberate pacing and shipped a Code of Conduct for Microsoft's MAI models, the first hyperscaler to publish a governance artifact alongside the stance. Techmeme | | Product | Salesforce added builder tooling to Agentforce. New Agentforce Builder, Agent Script, and Agentforce Voice capabilities aim to make agents faster to build and easier to control. Salesforce | | Infrastructure | Cloudflare's default handling of mixed-use AI crawlers changes today. Bots blending indexing with training and retrieval face new defaults on ad-supported pages, so teams relying on generic crawls should move to dedicated agents. Fast Crawl | | Research | A DeepMind study watched 100 agents self-organize. After one agent found a grading shortcut, most of the swarm sorted into problem-solvers and 24 that repurposed a bug-report tool to flag the issue to humans, a useful look at emergent oversight. MIT Technology Review | | Compute | Vera Rubin posted 7x Blackwell's tokens per megawatt. SemiAnalysis pre-release testing on DeepSeek V4 Pro shows Nvidia's Rubin NVL72 well ahead on energy-normalized throughput, good news for inference economics. SemiAnalysis | | Edge | MediaTek shipped a 2nm phone chip that runs 30B models on-device. The Dimensity 9600 Pro, built on TSMC's N2P process, brings agentic, on-device assistance to phones shipping this quarter. Reuters | | Policy | A new AI policy center launched under Jay Carney. Ron Conway's SV Angel is backing Project Blueprint, a bridge between labs and lawmakers on frontier governance, with public support from Altman, Amodei, and Hassabis. Crypto Briefing | | Security | Agent security is becoming its own discipline. GreyNoise detailed an attacker using AI agents to exploit PaperCut flaws at scale, underscoring why agent-aware defense and fast patch cadence now belong on the roadmap. Help Net Security |
| Quick HitsThe funding and deals board, in one line each. | Euclyd, based in the Netherlands, raised more than 200 million euros (about 230 million dollars) in a Series A co-led by Samsung and EQT's Scaleup Europe Fund to build a non-GPU inference chip, with ex-ASML chief Peter Wennink as chairman.CNBC | | Exein, an Italian embedded-security firm, raised 270 million dollars led by Headline at a 1.7 billion dollar valuation to secure physical AI at the chip level, with its runtime already shipping in over two billion chips.Financial Times | | Z.AI, formerly Zhipu, launched a roughly 5 billion dollar combined share-and-convertible-bond raise, with 60 percent earmarked for next-generation GLM foundation models and infrastructure.TechNode | | Firmus, an Nvidia-backed Australian AI data-center provider, is targeting an ASX listing to raise up to roughly 5 billion dollars for a 1.6-gigawatt compute buildout.Data Center Dynamics | | Kioxia is weighing a US ADR listing to raise at least 10 billion dollars in spring 2027 to meet AI memory demand.Bloomberg | | Celero Communications raised 275 million dollars in Series C for coherent DSP technology aimed at AI infrastructure interconnects.Crescendo AI | | Gimlet Labs raised 300 million dollars in Series B to advance applied AI research and products.Crescendo AI | | Savvy Wealth closed a 100 million dollar Series C for its AI-native platform for independent financial advisors.Crescendo AI | | Inspiren raised 70 million dollars in Series C at a valuation above 500 million dollars for AI-powered senior-living operations.Crescendo AI | | CrowdStrike launched an AI Partner Specialization within its Accelerate program to help partners resell, build, and secure agents on the Falcon platform.AI Agent Store | | GMI Cloud is building a 500 million dollar, Nvidia-powered data center in Taiwan to add regional inference capacity.MarketScreener | | Shanghai AI Laboratory released Atria Dawn Preview, a 744-billion-parameter agentic mixture-of-experts model trained to ground tool use in executable environments.arXiv |
| The Number80% Of new enterprise apps now ship with an AI agent The share of enterprise applications shipped or updated in early 2026 that embed at least one AI agent, up from about a third in 2024, per compiled 2026 enterprise-adoption research. That is the number that makes this Dreamforce week matter. Agents are no longer a side experiment; they are becoming a standard feature of the software your teams already use, and the baseline you are building against. When four in five new enterprise apps arrive with an agent inside, the strategic question stops being whether to adopt and becomes how well you deploy: which workflows first, what data to ground them in, and how to measure the return. The direction is set. The advantage now goes to the teams that get good at execution. | Counter-SignalStrategyThe returns are real, but they are uneven. The optimistic read of this week is the right one, with one refinement worth holding: the value of agents is real but unevenly distributed, and treating every use case as equally ready is how good programs lose momentum. The same 2026 adoption research that shows strong average returns also shows wide variation by function, with some agents paying back in a few months and others taking far longer as data, integration, and change management catch up. That is not a reason for caution about the technology. It is a reason for discipline about sequencing. The practical move is to treat the agentic enterprise as a portfolio, not a switch you flip. Start where the payback is fastest and the data is cleanest, usually a well-bounded, high-volume workflow with a clear metric, and let those wins fund the harder deployments. Measure your own returns rather than borrowing a keynote's, and expand as the numbers earn it. Read this way, the uneven returns are not a warning against agents but a map for how to roll them out, which makes the promise of this week easier to capture, not harder. | From the FieldThis is the most interesting the enterprise-AI market has been. Be in it. Every September the enterprise-software world gathers in San Francisco to decide where the future is going, and some years the answer is clearer than others. This is one of the clear years. The agentic enterprise is moving from a phrase on a slide to the way software is actually built, and the fact that Salesforce, Anthropic, and OpenAI are all telling the same story from the same stage is a signal worth trusting. When the platform, the model, and the runtime converge, building something real gets easier, and that is genuinely good news for the people who have to ship. So watch this week with curiosity rather than caution. The interesting question is no longer whether agents belong in the enterprise; adoption already answered that. The interesting question is where you start and how fast you learn. The teams that will look smart a year from now are the ones running a real pilot this quarter, grounding it in their own data, measuring their own returns, and letting early wins fund the next step. Take the keynote for the direction and the demos for the possibility, then go do the practical part. Pick one workflow with a clean payback. Put sensible guardrails around it. Ship it, measure it, and expand from evidence. This is the most interesting the enterprise-AI market has been, and the best way to enjoy an interesting moment is to be in it. Let's get to production,
AK | | The Agentic Enterprise Know more about AI than 95% of your peers. By 7 AM. A daily AI intelligence briefing for enterprise leaders, published by Spearhead. We build AI systems that work. Strategy. Engineering. Production. Outcomes. © 2026 Spearhead. All rights reserved. |
|