September 2026 Summaries

Rollouts and Security Reviewer are new software development bots designed to automate post-PR tasks that can delay safe production releases, including deployment monitoring, regression response, and security review. Rollouts connects source control, deployment systems, and telemetry tools to create editable monitoring plans before merge, compare production behavior with pre-deployment baselines, identify suspected regressions, and optionally notify authors, pause rollouts, or prepare revert pull requests. It is intended to detect localized issues, distinguish expected changes from failures, and identify missing instrumentation, with future support planned for feature flags, release trains, and deployment freezes. Security Reviewer evaluates every pull request in the context of the broader codebase, tracing data flows to identify vulnerabilities such as injection risks, broken authorization, exposed secrets, unsafe deserialization, vulnerable dependencies, and insecure configurations, while providing severity assessments, attack paths, and proposed one-click fixes. Both bots are available through the automations tab for Teams and Enterprise plans.

Cursor reports reducing agent token costs by 7% without lowering quality by optimizing its agent harness, which controls request assembly, context reuse, and subagent delegation. Key changes included cutting roughly 66% of the system prompt as models became more capable, dynamically loading infrequently used built-in and MCP tool definitions rather than including them in every request, and retaining only commonly needed tools in static context, reducing their description tokens by 60%. The company also improved prompt-cache reuse by placing explicit cache boundaries around stable instructions and tools, lowering cold cache misses by 20%, and reduced file-reading overhead by numbering only every tenth line while preserving citation usefulness. For longer tasks, Cursor adjusted subagent guidance and model selection to gain the context-isolation benefits of delegation while limiting duplicated work and unnecessary coordination. It plans to continue measuring context growth and applying similar harness-efficiency methods to Grok Bot.

Grab reports that Cursor has become widely used across its Southeast Asian technology organization, with about 98% of tech employees using it monthly and 75% weekly, supporting work in engineering as well as design, operations, finance, analytics, product management, and executive teams. An analysis of more than 100,000 sanitized messages from roughly 4,000 users found bug fixing among the most common activities across professional groups, while designers used the tool heavily for styling and layout and, after Git training, merged hundreds of reviewed UI fixes. Among engineers, frequent users handled a greater share of complex tasks, with Cursor contributing to more than one-third of merge requests and approximately half of suggestions being accepted; Grab says this has shortened some tasks from days to hours and enabled work such as refactors and tests that might otherwise be deferred. Non-engineering teams used Cursor for SQL transformations, deployment workflows, documentation, prototypes, and internally built tools, supported by workshops that trained several hundred people in five countries. Grab characterizes the adoption as an enablement effort rather than a mandate, emphasizing that AI tools allowed employees to independently complete work that previously required engineering support.

Cursor has launched Projects, a beta feature designed to manage long-running software work such as features, migrations, full applications, and ongoing maintenance through a coordinator agent that delegates tasks to potentially thousands of subagents. Projects run primarily in the cloud while using local agents when machine-specific testing is required, retain shared context and artifacts across agents over time, and can react autonomously to signals such as Slack messages, schedules, pull requests, and CI failures. Cursor reports that internal teams have used Projects for large migrations, design-system consistency, and the development of Projects itself, with new users merging 30% more PRs and users relying primarily on Projects merging six times as many. Common uses include researching, planning, implementing, testing, monitoring, and maintaining features; progressively applying migration patterns across hundreds of PRs; and “gardening” tasks such as detecting regressions, enforcing standards, and responding to bug reports. Available through Cursor’s navigation, Projects is intended for work extending beyond a single chat and is rolling out to all users.

Basis develops AI agents for accounting teams that perform complex, long-running tasks such as month-end close, tax returns, and audit work, producing outputs intended for professional review. Because these workflows involve hundreds of interdependent decisions and are difficult to assess solely by final outcomes, the company treats agent context—including prompts, tools, instructions, skills, and examples—as a production-critical input that must be carefully inspected and revised. Basis uses behavior specifications, written as Markdown documents separate from agent prompts, to define observable standards for recurring agent behaviors and enable judges to assess recorded work trajectories as true, false, or not applicable. Engineers use Cursor to write and refine both these specifications and the runtime context, while Braintrust evaluates whether agents followed the specified behaviors in production; identified failures then guide updates to tools, prompts, or execution frameworks. Basis reports that its agents can complete a Form 1065 partnership return in roughly six to seven hours compared with an estimated 30 to 40 hours of human work, and it has released its behavior-specification approach with Braintrust as an open standard for evaluating agent processes.

Cursor has introduced Self-Hosted Machines for Cloud Agents, allowing teams to run agent tool execution on dynamically scaled infrastructure within their own networks while Cursor continues to manage planning, inference, and agent orchestration in its cloud. The capability is intended for organizations needing direct access to internal services and repositories, specialized hardware such as GPUs or Macs, or custom operating systems and build pipelines, while preserving outbound-only HTTPS connections so Cursor does not initiate access into customer networks. Individual machines can be connected for personal use, while worker pools use queues, team-provided spawn scripts, idle timeouts, and optional hibernation to scale capacity, preserve workspaces, and serve tasks from multiple repositories. Cursor supports integrations with sandbox and compute providers including AWS Lambda, Cloudflare, Coder, Daytona, E2B, Modal, Namespace, and Vercel, and Linux workers now support browser-based computer use alongside Macs. Cursor-hosted environments remain the default, but the new option gives enterprises greater control over where agents run and what systems they can access.

Nokia’s Core Networks division is using Cursor’s AI-assisted, multi-agent capabilities to modernize workflows supporting highly regulated 5G mobile voice and data services that require five-nines reliability. Two engineers analyzed a hybrid, multi-repository codebase exceeding 50 million lines in two weeks, producing an evidence-based plan to break down a monolithic architecture into distributed services while reportedly avoiding months of work by more than a dozen specialists. Nokia also used Cursor to create a project-management tool in under a week that automated and visualized about 80% of a complex 5G function deployment workflow, and it has reduced software-defect root-cause investigations from weeks to days by analyzing logs, tickets, code, and debugging traces in parallel. The company is extending the technology across its software development lifecycle for requirements gathering, product requirements documents, architectural planning, and future automated cataloging of network elements, with the goal of pairing specialized AI agents with employees across development and delivery tasks.