Public transportation is one of the most complex operations on the planet. Every day, planners, schedulers, and operators make thousands of decisions; balancing driver availability, vehicle constraints, union rules, passenger demand, and real-time disruptions, all at once.
AI has promised to help with this complexity for years. Now, with the launch of Optibus Agent, it actually does.
Optibus Agent is an AI agent built for public transportation that understands transit's complex operational and compliance logic, personalizes to your organization and users, analyzes trade-offs and explains its reasoning, and acts, executing tasks like driver preferences and allocation directly in the platform, not just recommending them.
Most people are familiar with AI assistants: you ask a question, you get an answer. That's useful. But running a transit network requires more than answers - it requires action.
An AI assistant answers questions. An AI agent performs tasks.
Optibus Agent is built to do both. It understands your question, works within your data, and can execute tasks directly inside the Optibus platform, and it doesn't just explain what to do. It does it, with your review before anything is applied.
"When we built the Optibus Agent, we looked at how agents and AI can help with the complexity of running a public transit system."
— Amos Haggiag, CEO & Cofounder, Optibus
The difference matters. An assistant can tell you how to plan a trip to Rome. An agent books the flights, the hotel, and the attractions. Applied to public transit, which is far more complex, that distinction changes everything. Tasks that once took days of manual work across network changes, timetables, duty scheduling, vehicle assignments, and passenger notifications can now be directed to an agent.
With the official launch of Optibus Agent at our webinar on transforming public transportation with agentic AI, we’ve released initial use cases, with more capabilities following quickly. These are already delivering real value:
Importing data from external sources is often a back-and-forth of uploading, failing, and trying to fix. The agent takes the error messages and the underlying file, diagnoses what's wrong, corrects it, and lets you download the fixed file, saving hours of work. When it encounters something it can't resolve from context alone, it surfaces exactly what's missing and asks you, rather than guessing or making up data.
Setting up driver preferences, union rules, and break logic used to require deep platform expertise. The agent knows the platform inside out and can translate what you describe in plain language, like a complex accumulated break rule across multiple rest and meal break types, into the correct configuration in the system. It asks clarifying questions where needed to make sure it understands your intent before acting.
Instead of manually filtering across multiple data slices to find an available driver, you ask the agent directly. It searches within your data and surfaces the right person instantly, including querying audit logs to understand the history and reasoning behind past decisions, for users with the appropriate permissions.
When buses run early or late, the agent has access to real-time vehicle data, identifies which buses are off-schedule, and helps controllers take action fast, drafting messages to drivers and sending them through the platform's APIs. The controller validates and sends; the agent does the legwork.
The agent also functions as an always-available support resource, answering how-to questions by drawing on Optibus's full knowledge base and giving you a contextual answer without needing to search through manuals or contact support.
Below, we've rounded up the questions our audience asked live during the webinar, along with our answers.
A validator tells you what's wrong. The agent diagnoses why it's wrong, fixes what it can based on the file's context and GTFS specs, and for anything it can't resolve on its own, it tells you exactly what information is missing and asks you directly. It will never make up data or guess. When the answer requires human input, the human is asked. That interactive, context-aware approach is what sets it apart from a one-off validation check.
It doesn't, and that's intentional. The agent is an interface to Optibus's underlying optimization engine, not a replacement for it. The algorithms that power scheduling and planning are the same ones we've been developing for over a decade. What the agent does is translate your intent into the right rules, preferences, and configurations so that the engine can do its job. You describe what you want to try; the agent sets it up correctly; the optimization runs. You evaluate the result and decide whether it's the right direction. The agent is not making optimization decisions - you are.
Yes. In disruption scenarios, for example, the Optibus agent doesn't just identify which buses are running off-schedule. It proposes what to do about it, whether that's drafting a message to a driver, flagging an issue to a controller, or taking action through the platform. Every recommendation comes with a human confirmation step. The agent proposes; you decide.
We use multiple best-in-class models, including Gemini, OpenAI, and Anthropic, and we select between them based on continuous internal benchmarking for each specific type of task. We don't train our own large language models. What we've built is the layer on top: the transit-specific knowledge, the platform expertise, and the controls that make sure the right model is used in the right way for each interaction.
We measure above 95% accuracy across agent interactions. A key part of how we achieve this is context discipline; the agent only has access to the data relevant to the specific task at hand, which prevents it from pulling in unrelated information and hallucinating. We also use AI-as-judge evaluation, where a second model reviews responses before they reach users. Accuracy improves over time through the feedback mechanism built into the platform: every agent response has a thumbs up/down, and that feedback directly informs how we improve the agent's instructions. That said, this is generative AI, and human review remains important, which is why every proposed action requires your confirmation.
Yes. Optibus has a control and dispatch product where you can see real-time delays and incidents, communicate with vehicles, and make changes on the fly. The agents work within that product too, with access to live vehicle data, the ability to identify which buses are running off-schedule, and the ability to draft and send messages to drivers through the platform. Optibus also offers its own AVL hardware if needed, or can connect to your existing AVL system.
The Optibus platform as a whole, with or without agents, is designed to reduce operational costs, cut dead mileage, and reduce overtime. On the revenue side, it improves service performance, increases frequency reliability, and helps operators win tenders and franchising bids. Agents are integrated across all of these areas: planning agents are more focused on matching service to demand and improving performance; scheduling and operations agents are more focused on cost optimization. As agents become more capable, the ability to factor in revenue and demand data more directly into scheduling decisions is a direction we're actively developing.
This is exactly the kind of work agents are designed for. Tasks that are manual, repetitive, and time-sensitive are high on our priority list. In the near future, agents will take on more and more of this type of work, reducing the manual burden on your team. Over time, as agents become more capable, they'll also elevate less experienced team members, giving them access to the same expertise that today requires years of platform knowledge to build.
No. We don't use your agency's operational data to train any models, public, shared, or internal. We use public best-in-class models and wrap them with our own transit knowledge and platform logic. What we do track, like any SaaS platform, is how people use the product, to understand where users are struggling and improve the experience. You can also choose to give direct feedback on agent responses, which helps us improve the instruction set. Neither of these involves your operational data being used for model training.
The agents shown at launch on June 17, 2026, and additional capabilities being released in the coming weeks, are included in your existing Optibus license at no additional cost. As we develop more advanced agents with larger operational impact, those will be priced separately, but always with a clear ROI case behind them. The goal is that the value delivered justifies the investment, in the same way the core Optibus platform does today.
Accuracy, transparency, and keeping humans in the loop are foundational to how Optibus Agent was built. Every proposed action requires your confirmation. Every response can be rated. Every piece of feedback makes the agent better.
We're building this to work the way your team works: supporting the judgment of experienced transit professionals, not trying to replace it.
What we've launched is the foundation. We're moving from tools that perform tasks to a team of experts that assist with decisions and take action. In the near term, agents will handle more of the manual work. Over time, they'll shift to anticipating problems before they happen, proactively alerting planners and operators rather than waiting to be asked. And eventually, Optibus agents will connect to the broader ecosystem of systems your agency already uses, becoming the connective intelligence across your entire operation.
The end goal hasn't changed. Technology is great, but what we're all here for is better public transportation; more efficient, better matched to demand, and more reliable for the passengers who depend on it every day.
We hosted a live webinar where our co-founders CEO Amos Haggiag and CTO Eitan Yanovsky walked through live demos of Optibus Agent across scheduling, operations, and disruption management.
Further Reading