AI agents have seen widespread rollouts across the enterprise, but friction remains. Today, Salesforce announced that it will be updating Agentforce, releasing a family of “job-ready agents,” purpose-built to perform long-horizon tasks across sales, service and commerce workflows.

The newly announced portfolio of agents includes: Piper, an inbound pipe gen agent that works across websites and inboxes to qualify and convert leads, Hunter, an outbound sales agent that collaborates with sales reps to support the sales pipeline from research to outreach, Casey, a help agent that resolves customer service issues across voice, SMS, WhatsApp and web chat. Other agents automate IT and HR requests and product comparisons. Salesforce’s move to offer purpose-built agents as part of Agentforce represents an attempt by the organization to make AI agents more accessible to end user organizations. Agent deployment has remained a struggle for enterprises, with Deloitte noting that only 11% of organizations have agents in production, despite 38% piloting them. Likewise, Gartner expects that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls.

The news comes less than a week ahead of Dreamforce, Salesforce’s annual technology conference, which includes 1,600 sessions with speakers including Marc Benioff, Omar Sultan Al Olama, Dario Amodei, Roland Busch, Boris Cherny, Terry Crews, Matthew McConaughey, Reese Witherspoon, Sterling Brown and more.

Throughout the AI race, organizations have experienced challenges in gaining value from generative AI and LLMs. One of the most widely cited studies, conducted by MIT and released in 2025, found that 95% of AI pilots fail to generate a noticeable ROI. While the study’s methodology has been criticized , the studies presented by Deloitte and Gartner above highlight the difficulties in deploying agents in the enterprise.

For Jayesh Govindarajan, EVP of Salesforce AI and head of engineering, part of the issue comes down to the limitations of LLMs. “The main trend that we are now seeing play out that we are fairly early to with Agentforce and generally agentic AI, is that LLMs by themselves are not enough,” Govindarajan told me in a video interview. “You need a harness around it that is able to bring in the right context, that is able to bring in the control, and has access to a bunch of tools to go get the job done.”

“There is very clear need for long-running, long-horizon agents. There’s very clear need for self-optimization of agents. It’s hard to manage them, and there’s a very clear need that’s emerging on these task-specific, out-of-the-box, easy to use, easy to set up agents," Govindarajan said.

Govindarajan notes that building an autonomous system and having it run at scale is not a trivial task. He also says that the enterprise is going in the direction of giving agents tasks that are not single-shot tasks, but tasks working toward a higher level goal, which the agent runs in a loop until completing its objective.

Long horizon agents have the potential to be significantly impactful with less human intervention. In fact, according to data provided by the company, the Piper pipeline agent has surfaced $82.5 million in leads for Salesforce since its launch in April, while the organization claims to have managed over 1 million agent conversations on the Salesforce website.

Salesforce also shared that a number of customer organizations had adopted its long-horizon agents, with 50% of travel platform Engine’s chat inquiries being “fully resolved” by their help agent Eva, 60% of travel and spend management provider Perk’s sales pipeline being built by the outbound sales agent Hunter and 70% of Autism Queensland’s administrative requests being resolved by an employee service agent.

More broadly, the news highlights the growing demand for agents that can perform long-horizon tasks in the enterprise over a prolonged period of time. At the same time, agents’ capabilities are improving, with a study released in July by METR finding that the frontier AI time horizon, or the length of tasks AI agents have been able to complete autonomously with 50% reliability, has doubled approximately every seven months since 2019.

As of 2026, long-horizon agents are picking up significant interest, with venture capital firm Sequoia going as far as to say “long horizon agents are functionally AGI.” OpenAI has also identified a trend toward long horizon tasks, with a recent blog post stating the startup is moving toward developing an automated AI researcher by March 2028, with qualitative impressions and internal data indicating that delegation of higher-level and longer-horizon tasks is becoming more common in coding agent workflows.

Rebecca Wettemann, CEO and Principal Analyst of technology analyst firm Valoir, argues that long-horizon runtime is becoming “table stakes” for AI platforms like Salesforce, Genesys and ServiceNow that want to be able to support and control the orchestration of more complex tasks and agents.

“Long-horizon tasks are important because real work doesn’t happen in one shot,” Wettemann told me via email. “To really delegate something to an AI agent, it needs to be able to persist, track state, recover errors and reengage a process without human prompting. The business case for agentic AI largely rests on autonomous task completion, and the potential benefit is simply much greater for multi-turn agents that can eliminate hours of human follow-up than a simple one-turn agent.”

Salesforce’s additions to Agentforce also highlight how taking a narrow approach to automating workflows can potentially reduce friction. “Prebuilt, pretested agents reduce the time and expertise needed to get agents up and running, and reduce the cost and risk to manage them over time. As organizations move from AI FOMO to FOMU (fear of messing up), Salesforce recognises that it can’t send a forward-deployed engineer into every account so it needs more prebuilt agents to drive adoption,” Wettemann said.

“Prebuilt agents mean companies don’t have to reinvent the wheel. Most don’t have the team to build agent workflows from scratch, so a ready-made library is what gets this into production. This is a head-to-head fight with ServiceNow for the front office. The more open platform wins this one,” Shashi Bellamkonda, principal analyst at technology consulting firm Info-Tech Research Group, told me via email.

The approach also opens the door to continuous optimization. Alongside this new family of agents, Salesforce announced Agent Optimizer, which Govindarajan describes as an agent that works alongside customer-built agents. He says its sole task is to help build and refine the agent, to identify where it fails and spot what to fix. “We are starting to get into the realm of self-improvement,” Govindarajan said.

In this sense, using agents to automate narrow processes, can help to provide opportunities for continuous improvement long term, as errors in simple, repeatable processes are identified and mitigated over time.