Optibus

Everything the agent reads and writes today sits inside the Optibus platform. It sets driver preferences and allocations directly in the platform, returns corrected import files, and drafts messages to drivers and "sends them through the platform's APIs".

Reach beyond the platform is stated only as intention: the recap reads "eventually, Optibus agents will connect to the broader ecosystem of systems your agency already uses", which is a plan, not a shipped capability. The Agent page confirms the boundary, "runs entirely within the Optibus platform", and there is no agent-side tool surface for custom tools. The Operations API's HR and fleet sync guides and outbound webhooks are platform integration a customer builds, and connecting the platform to an existing AVL system is the platform's integration, not the agent's.

Sourceblog.optibus.com webinar recap (6 July 2026), optibus.com/product/agent; readread 2026-09-15

A planner's intent becomes configuration for Optibus's deterministic optimization engine, and the agent carries out multi step operational fixes.

The webinar recap documents the mix of agent steps and deterministic steps as the product's design: the agent "is an interface to Optibus's underlying optimization engine", translating a planner's intent into rules, preferences and configurations, after which the deterministic optimization runs and the person evaluates the result.

Multi-step execution runs across live use cases: a data import is diagnosed, corrected and returned as a fixed file, falling back to asking the user when context cannot resolve it; and disruptions are identified from live vehicle data, a remedy proposed, and a driver message drafted and sent through the platform.

The orchestration is a vendor-designed pattern the buyer directs by intent rather than a workflow the buyer authors, and loops and retries, versioning and reuse across teams are not documented.

Sourceblog.optibus.com webinar recap (6 July 2026), optibus.com/product/agent; readread 2026-09-15

Live operational data in the platform and Optibus's own knowledge base are what the agent works from; new data is usable at once and nothing is trained on it.

The Agent page says it "runs inside the Optibus platform, connected to your real data" and understands "schedule KPIs, audit trails, duty assignments, and real-time vehicle punctuality"; the webinar recap has it search the customer's data to find eligible drivers, read real-time vehicle data during disruptions, and answer how-to questions from Optibus's full knowledge base.

Customer data enters the platform and is queried at run time, so new knowledge is usable without retraining, and the recap states "we don't use your agency's operational data to train any models."

Access is scoped: context discipline limits the agent to data relevant to the task, and audit-log queries are limited to users with the appropriate permissions. Optibus draws on more than a decade of public transportation deployments, but says it does not train its own models. Whether answers cite their sources is not documented.

Sourceoptibus.com/product/agent (modified 31 July 2026), blog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

Every action the agent proposes needs a person's confirmation before the change is applied.

The webinar recap describes the agent executing "tasks directly in the platform, with your review before anything is applied"; in disruption management "the controller validates and sends" while the agent drafts; and the FAQ states that "every proposed action requires your confirmation" and that every recommendation "comes with a human confirmation step".

The confirmation is a step in Optibus's own product before a change is applied, not an obligation placed on the customer. The agent also escalates to the person when it cannot resolve something from context rather than guessing, and Optibus describes the planner as always in the driver's seat.

What is not documented is any difference in autonomy between low and high risk actions, or routing of approvals into existing ticketing channels.

Sourceblog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

ISO 27001:2022 certification comes with role based access, single sign on with SAML and SCIM provisioning, and multi factor authentication. Every optibus.com footer links "ISO 27001:2022 certified" to the certificate, and the Security & Trust page carries a badge strip for NIS2, Cyber Essentials, GDPR, CSA STAR Level One, ENS and ISO 27001.

The Identity & access tile names role-based access control, single sign-on with SAML and SCIM, and MFA, which answers which identity providers and provisioning standards are supported. Access scoping reaches the agent itself: per the webinar recap, the agent queries audit logs only "for users with the appropriate permissions".

Sourceoptibus.com/company/security (modified 31 July 2026), optibus.com footer, blog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

Traceability of every agent action is stated, though the audit surface itself is not shown. The Agent page, under "Full audibility and accountability", states "every action the agent takes is traceable", and that the agent runs entirely within the platform "with full auditability".

That covers audit events and execution history for the agent's own conduct in production, and every agent response carries a thumbs up or down rating. An agent that explains its reasoning and answers "why does this schedule carry more overtime" is explaining the optimization engine's schedule, not making the agent observable.

Step by step inspection of prompts and tool calls, export to monitoring or SIEM, and retention by plan are not documented. The platform audit logs the agent queries for users, such as who approved an absence, are team activity the agent reads as data, not a record of the agent's own behavior.

Sourceoptibus.com/product/agent (modified 31 July 2026); readread 2026-09-15

Memory of any kind is undocumented: no session, conversation, workflow or long term memory is described, and no persistent store.

The agent is said to personalize to the organization and its users and to improve through a feedback loop that keeps refining its instructions, but the feedback loop is thumbs up or down on responses that informs how Optibus improves the agent's instructions, which is learning absorbed on the vendor's side, and "personalizes" names an outcome with no mechanism. Context discipline, the agent seeing only data relevant to the task, scopes retrieval rather than persisting anything.

Sourceblog.optibus.com webinar recap (6 July 2026), optibus.com/product/agent; readread 2026-09-15

The platform runs as a single cloud service, and whether data can be pinned to a region is unknown because the document most likely to say could not be read. The Security page and the Agent page describe one cloud-native platform, and the agent "runs entirely within the Optibus platform"; no VPC, dedicated or self-hosted option is documented. The Security page's "multi-region redundancy" is resilience, not residency, and names no region.

For a single mode cloud product the open question is whether model traffic and data storage can be pinned to specific regions, and the site does not answer it. The Sub-processors List linked from every footer is closed to automated retrieval by the site's robots rules, and it is the first party document most likely to name hosting locations. The Operations API issues an account-specific base URL per customer, which suggests per-tenant hosting but names no region.

Sourceoptibus.com/company/security (modified 31 July 2026), optibus.com/product/agent, apidocs.ops.optibus.co; readread 2026-09-15

One embedded agent with capabilities shaped around each role ships to every customer, not a set of separate agents a buyer chooses among. One Optibus Agent is "included for all customers at no additional cost", with capabilities shaped around each role (planners, schedulers, operations managers, control dispatchers), and described in the recap as planning, scheduling and operations agents.

Nothing is selected, cloned or started from: every capability is switched on inside the one embedded agent, and removing the scheduling capability would leave a gap in the same agent rather than a separate whole product. There are no templates to judge as demos or production-ready starting points. A named agent set including Preference Designer has been described, with further agents such as Schedule Analysis announced as on the way but not yet shipped.

Sourceoptibus.com/product/agent, blog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

Conversation reaches the agent at every stage of the platform, but it only starts when someone asks; nothing wakes it on its own. The conversational surface is real and unusually broad, embedded across the platform, with the Agent page reading "One platform. One conversation. Real action." Planners, schedulers and control dispatchers each ask in plain language and get action across planning, scheduling, real-time dispatch and service evaluation, which answers where the agent works.

A conversational trigger is the person doing the waking, so an agent that acts when asked, rather than starting on an event, has no autonomous trigger. Proactive alerting appears in the recap only as a future capability, and the Operations API's outbound webhooks announce platform changes to subscribers rather than waking the agent.

Sourceoptibus.com/product/agentread 2026-09-15

Several named models, Gemini, OpenAI and Anthropic, sit behind the agent, and Optibus picks among them per task. The webinar recap FAQ, "What AI models does Optibus Agent use?", states it uses multiple models including Gemini, OpenAI and Anthropic and "selects between them based on continuous internal benchmarking for each specific type of task", and that Optibus does not train its own large language models. That is disclosed vendor-side routing across more than one named model.

Optibus is the one choosing: no customer or admin selection, workspace model policy or bring your own key option is documented.

Sourceblog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

Optibus publishes a full, versioned Operations API reference on its own domain, with read-write endpoints, webhooks and integration guides.

The reference at apidocs.ops.optibus.co is the Optibus Operations API (version 2.11.2), a Redocly reference with a downloadable OpenAPI specification, API-key authentication, and a documented semantic-versioning and deprecation policy: breaking changes ship as new versioned endpoints with at least three months' notice.

It has read write endpoints across drivers, absences, vehicles, rosters, operational plans, payroll and statistics, outbound operational event webhooks, and step-by-step integration guides for external HR and fleet management systems. The versioning policy speaks to stability; no SDK is named.

Credentials and the account-specific base URL come from the customer's Optibus Customer Success Manager, which is ordinary enterprise provisioning of a publicly documented API. The HR and fleet sync is platform integration a customer builds rather than agent tool calling, and the outbound webhooks announce platform changes rather than waking the agent. No MCP surface is documented.

Sourceapidocs.ops.optibus.co (v2.11.2); readread 2026-09-15

A second model checks the agent's responses before users see them, and accuracy is reported above 95%, but buyers get no readable test results or pre release harness. The webinar recap describes "AI-as-judge evaluation, where a second model reviews responses before they reach users", a thumbs up or down on every response that informs how Optibus improves the agent's instructions, and a measured accuracy "above 95% across agent interactions". That is a live quality gate plus a vendor-run improvement loop.

None of it is readable by the buyer: the judge's verdicts are not exposed, the 95 percent figure is a vendor wide claim rather than a result a customer can read for its own agents, and nothing shows workflows tested against datasets before production.

Sourceblog.optibus.com webinar recap (6 July 2026); readread 2026-09-15

Inside its own platform the agent has direct access to the data, so it never operates other software through a screen. The Agent page describes it finding a vehicle or driver and then it "navigates to it, and acts", and it "runs entirely within the Optibus platform", navigating the platform interface, pulling the relevant configurations and preparing the groundwork for review.

That is in-product interface operation rather than general browser or computer use. An agent moving through its own vendor's product, with direct access to the data and actions underneath, is not driving software it has no API for, and breakage when UI elements change is not a concern when the vendor owns the interface.

Sourceoptibus.com/product/agent (modified 31 July 2026); engineering post read as a search result only, not graded; readread 2026-09-15