Cisco and Kore.ai launch tools to manage enterprise AI agents

Cisco said Dialog assigns agents to keep consumers happy and to “keep working on the customer’s behalf after a conversation ends,” Engadget reported. Under Dialog, agents can coordinate across people, agents and backend systems. Cisco also said the platform speeds onboarding by treating AI agents like new teammates, allowing them to review existing employee handbooks, knowledge bases and operating procedures. Cisco promises a safety-forward experience through a collaboration with data analysis company Splunk.

The updated Webex will let anyone in a workspace directly invite AI agents into spaces, meetings and calls to execute complex, multi-step work across the platform and in third-party applications. Cisco gave an example of an agent quickly pulling up statistics during a presentation and updating a PowerPoint on the fly. Engadget noted that multiple studies show AI agents remain prone to mistakes that lead to cascading failures, with failure rates of 60 to 70 percent on multi-step tasks. It said a single mistake in one step can throw off an entire process, and an agent with 95 percent per-step reliability would succeed only around 36 percent of the time in a 20-step workflow. The report also cited Carnegie Mellon University researchers, who found earlier this year that the best-performing AI agent at the time, Google’s Gemini 2.5 Pro, failed to complete real-world office tasks 70 percent of the time.

Kore.ai said Autoloop is designed to keep tuning agents after they go live. Businesses set targets covering task completion, adherence to business rules, cost and accuracy, with accuracy judged on whether answers are backed by the enterprise’s own data, according to SiliconANGLE. The engine optimizes for all goals at once, and every proposed change is scored against the full set, so lower token spending cannot quietly come at the cost of safety or an agent’s ability to finish a job.

Autoloop writes the first version of each agent, tests included, from a company’s existing operating procedures, then iterates until every goal is met. Once an agent is in production, real interactions start fresh rounds of optimization. Kore.ai’s StateTrace evaluation layer spots where a goal slipped by judging agents against the full record of what they did in production, down to each handoff, tool call and state change across a network of agents. A five-layer validation architecture makes most checks deterministic, which Kore.ai says keeps round-the-clock optimization affordable at enterprise scale.

Kore.ai’s Agent Blueprint Language, introduced earlier this year, compiles routing, business rules and guardrails into an executable state machine, so each step in a trace maps back to a specific part of the blueprint. Autoloop can then rewrite only the piece responsible for a miss. “You can’t optimize what you can’t see, or fix precisely what you can’t express precisely,” said Prasanna Arikala, Kore.ai’s chief technology officer and chief product officer. He credited the two technologies with making automatic optimization practical.

Kore.ai said its own software development reflects the approach. About 6,500 commits a month to its 2.6 million-line production codebase now come from AI agents working under 68 always-on guardrails. Founder and Chief Executive Raj Koneru said companies that manage to scale AI “will be the ones using AI to build, govern and optimize AI.” Autoloop is available now to all customers on the Kore.ai Agent Platform’s Artemis edition, which launched in May. Kore.ai counts more than 500 Global 2000 organizations as customers.