Optibus ships an AI agent to run day-to-day public transit operations

What changed: Optibus unveiled what it describes as a platform-embedded AI agent designed specifically to help transit agencies with planning, vehicle and driver scheduling, dispatch, and real-time operational decisions. The company frames it as an AI agent purpose-built for public transportation workflows rather than generic back-office automation.

Why it matters: Transit operators can move from static schedules to dynamic, data-driven decisions about routes, staffing, and service changes without needing in-house AI teams. For city agencies facing budget and driver constraints, an embedded agent could cut manual spreadsheet work and speed up responses to disruptions like traffic or breakdowns.

Try/watch: If you run or serve a transit operation, map your current scheduling and dispatch tools against what Optibus's agent can automate, and pilot it on a single depot or route before scaling.

Databricks pitches its Lakehouse as the operating layer for agentic AI

What changed: A wrap-up from Databricks' Data + AI Summit 2026 argues that the Databricks Lakehouse is evolving into the operating layer for agentic AI, combining data, compute, and governance to run autonomous AI workflows end to end. The analysis highlights how Databricks is framing its platform as the place where organizations orchestrate agents that can access data, call tools, and trigger downstream systems.

Why it matters: Many teams are experimenting with agents in sandboxes but struggle to connect them safely to production data and pipelines. Standardizing those agents on a governed data platform reduces integration work and makes security and observability part of the default setup instead of custom glue code.

Try/watch: If you already use Databricks, inventory which existing jobs and notebooks could be wrapped as tools callable by agents, and push your data team to define guardrails for which tables and actions agents can reach.

Security experts flag identity and access as the weak link for agentic AI

What changed: A new security advisory outlines how agentic AI systems that can take actions on behalf of users introduce fresh identity and access management risks. The piece warns that when agents can sign in to applications, make configuration changes, or move data, they effectively become powerful machine identities that attackers can target or misuse.

Why it matters: Many teams are rushing to connect agents to internal tools without giving them the same least-privilege, credential hygiene, and monitoring controls used for human admins and service accounts. A compromised or misconfigured agent could exfiltrate sensitive data or trigger costly changes faster than a human because it operates at machine speed.

Try/watch: Treat each agent as its own identity: give it unique credentials, narrow roles to specific actions, log everything it does, and involve security early before connecting agents to critical systems.

Telecom brief: on-device agentic Gemini opens new automation plays for operators

What changed: An industry analysis explores how running Google's Gemini models directly on devices can enable agentic AI for telecom operators, with agents reasoning over local context while coordinating with network services. The article argues this shift lets telcos push more intelligence to the edge, from customer-support helpers on phones to agents optimizing network resources in near real time.

Why it matters: Operators have struggled to differentiate beyond connectivity, and on-device agents create new ways to personalize plans, reduce support load, and automate routine network operations without always calling back to the cloud. Local processing also reduces latency and can keep more customer data on the device, which helps with both user experience and regulatory pressure.

Try/watch: Network and product teams at operators should identify one or two high-friction customer or operations journeys—such as plan changes or outage triage—and prototype how an on-device agent could handle most of the steps.

Practical playbook: how to design memory for production AI agents in 2026

What changed: A new technical guide lays out patterns for building AI agent memory in 2026, covering how to combine short-term conversational context with longer-term records stored in databases or vector search. It emphasizes that reliable memory is now a design problem—choosing what to save, how to retrieve it efficiently, and when to forget—to keep agents both useful and affordable.

Why it matters: Many early agent pilots fail because context windows overflow, prompts become brittle, or costs spike as teams naively stuff every past interaction into each request. Engineering memory as its own subsystem makes agents more predictable, reduces token spend, and avoids privacy issues from retaining unnecessary data.

Try/watch: For any agent in production or pilot, diagram what information it needs across a week of use, then split that into ephemeral session context and durable memory, with explicit rules for what gets written where and when.