Enterprise AI agents can do the job, but the infrastructure that allows them to talk to each other, prove that they are to be trustedand should be checked when something still under construction goes wrong.
Here’s a look at how five startups are addressing this gap in orchestration, observability, connectivity, and security, as shown below. VB Transform 2026.
BAND manages all the agents you run in the background
In the very near future, agents will be deployed everywhere and they will work on our behalf, noted CTO and co-founder Vlad Luzin. BAND.
As he describes it: They will receive tasks, visit registries, recruit other agents to help them, assign subtasks to AI peers in the “conversation space,” aggregate and share the results, then return a summary to the human user.
To make this happen, BAND builds a coordination infrastructure layer for multi-agent AI systems.
Why doesn’t Telegram, Slack or Discord solve the problem? Luzin noted that these platforms were built for people. Agents must be started manually in multiple steps, and they cannot see each other; “They’re still alone in some kind of digital solitary confinement.”
Likewise, Claude is stateless, and developers often have multiple sessions open at once, switching between different tasks — something Luzin says creates real friction.
The challenge connects remote processes, which Luz calls the distributed system problem.
“The transport layer of how agents communicate in real-time needs to be addressed first,” he said. Conversations cannot occur via IPs and URLs; agents need to be mapped to an abstraction layer so they can talk across channels, conversation spaces, and platforms.
“Agents see each other. They understand each other. They can collaborate together. They discuss problems. They solve problems and ask for someone else’s review,” Luzin said.
According to Luzin, BAND supports autonomous workflows that can run for eight to 20 hours and are compatible with A2A and MCP protocols. Importantly, people can join the conversation as agents talk and discover each other, he said.
“We can record and display all the tasks your agent creates in real time,” Luzin said.
Conifers help defenders move at machine speed
The biggest challenge defenders face today is that they’re still running at human speed, but opponents are running at machine speed, said Tom Findling, CEO and co-founder. Conifers.
Attackers already accept agents, Findling said, and only need to succeed once to penetrate an enterprise. Malicious campaigns that used to take months and weeks now take hours or even minutes. On the other hand, security operations are fragmented, manual, inefficient and slow.
Findling said Conifers takes the various components of cyber defense — private intelligence, hunting, detection, engineering, investigation, response — and outsources them. They then broke down the silos between them, he said. Different agent systems can communicate with each other so that operational protection and active protection are always active and adaptive.
Findling said Conifers’ system compressed hold times from 7 hours to 12 minutes, and the company can turn around complex cyber investigations in four minutes or less.
He emphasized the importance of connecting to an enterprise’s existing security tools, whether it’s endpoint detection and response (EDR), security information and event management (SIEM), posture management, or others. Conifers help clients understand their security posture, pain points, what controls are working and what aren’t, and areas to invest in for the best ROI.
“The threat landscape is changing, detection remains the same and threat intelligence is not operationalized,” Findling said. “This is a job for agents.”
Raindrop AI agent generates an audit log
One of the defining challenges of the current era is finding critical problems in AI agents, says Ben Hylak, the company’s CTO. Raindrop AI.
This is what he calls the “double whammy”: as agents become more skilled, complexity and timelines increase; in some cases they run for hours or days. Second, problems in sectors such as health or defense are catastrophic.
“This problem is getting worse as models and agents improve,” Hylak said, “and I think there’s good reason to believe it’s going to get worse.”
Raindrop’s AI platform finds critical issues with agents in production and simulates fixes based on past user behavior, Hylak said. This allows teams to confirm that a fix works as intended before going live without introducing unexpected side effects.
The startup’s reinforcement learning (RL) platform optimizes trailers and builds models directly from Raindrop data, he said. Its pre-deployment simulation engine helps determine which fixes will actually affect production; its live A/B test shows these changes in action.
Messages, tool calls, retries and errors are captured together, and human users are notified (usually via Slack) when a problem occurs, he said. Models are built for each customer, and alerts drive continuous learning between models and trailers. “It’s condensed into something that’s actually navigable, understandable, easy to validate,” Hylak said.
Arcade gives agents the security clearance they need to take action
AI agents are designed to do all kinds of work for you, but they often hit three major hurdles: authorization, management, and reliability.
Agents need a new security architecture to act on behalf of real users with real permissions, said Sam Partee, the company’s co-founder and CTO. Arcade.dev.
Partee said his company’s secure agent runtime provides this level of authentication and authorization so agents can pass critical security checks. It also provides observability so that human users can watch everything the agent does. Actions are assigned to a precise moment in time with least privileged scope.
Arcade is available as an installable plugin and can be deployed locally in a cleanroom-like environment; Companies can continue to use their own access and security tools, Partee said. Anything running on Arcade is bound by the same role-based access controls (RBAC), intrusion detection and prevention systems (IDPS), policies, rights, and other pre-established checkpoints.
Partee noted that Arcade “addresses a very pervasive supply chain attack problem; it’s incredible.” Security and observability continued to be difficult because “basically, the abstraction was wrong.”
Omilia struggles "it’s not right" The CX challenge
Addressing enterprise customer experience (CX) is “really not easy,” said Claudio Rodrigues, CPO Homily.
Heuristic-based systems are manageable but slow; agent systems are fast but unpredictable, Rodrigues said. Omilia built its platform to deliver both control and speed together.
The agency’s self-learning offering is built on the philosophy of observing customer service transactions as they actually happen, not abstractly. Omilia’s agents observe issues firsthand, listen to every customer and agent interaction, capture data, API features, screen captures and standard operating procedures (SOPs), then map them to use cases for customer support, he said.
Rodrigues said contact centers should be a source of revenue, and Omilia’s differentiator is its speech-to-text systems, management and observability layers.
AI generates insights, suggests improvements, creates automatic conversational agents, pulls data from documents and APIs, and designs dialog flows. Human experts can then test real and simulated interactions and put them into production under their supervision. Rodrigues said Omilia combines all of these into an enterprise-wide engine that continuously learns over time.
Rodrigues said the company is responsible for more than 3 billion calls per year, more than 1 million voice calls per day in some deployments, and has seen a 30-45% improvement in time to resolution (TTR). Omilia agents say they generate 21 times more sales revenue than human agents.
In a mature deployment, automation “easily” reaches 80-90%, he said. However, “the human in the loop is still very fundamental to us.”
