Netlify

An agent run acts inside the customer's Netlify project and its GitHub repository: it edits files, builds a Deploy Preview, and can open or update a pull request, with deploy events sent to Slack, webhooks or email. Agent Runners document no catalog of connectors or custom tools the agent can call into other systems. The customer's own function code can call anything, but that is an integration the customer writes. Separately, the Netlify MCP server lets outside agents act on Netlify, and the AI Gateway connects the customer's code to models.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/make-changes-with-agent-runnersread 2026-09-22

Async Workloads, part of Netlify's AI runtime, give the customer's code a durable workflow runtime: a workload is broken into discrete steps with step.run, failed steps retry with default or custom limits without repeating steps that succeeded, steps can sleep or be scheduled, and workloads start on events.

A step can call a model through the AI Gateway between deterministic steps, so deterministic work and agent steps mix in one workload, with retries built into the runtime. Branching is written in code rather than drawn in a visual flow. The task sequence inside Agent Runners belongs to Netlify, not the customer.

Sourcedocs.netlify.com/build/async-workloads/multi-step-workloadsread 2026-09-22

Agent Runners ground each run in the customer's own project: the agent starts with the code, environment variables, build settings, deployment pipeline and Netlify primitives, and a team can add project context, such as prompt guidelines or a link to public documentation, that applies to every run. That is company data the agent works from.

No retrieval structure that Netlify maintains over the customer's content, such as an index or a search a query returns, is documented; the agent reads the project it is working in. Netlify Skills are reference material about Netlify itself, not the customer's corpus.

Sourcenetlify.com/platform/agent-runnersread 2026-09-22

Nothing an agent run does reaches production until a person ships it: every Build mode run produces a Deploy Preview and a list of changed files for review, and shipping means merging the pull request or pressing Publish, which requires publishing permission; Internal Builders can propose changes while Developers control deployment, and any change rolls back in one click. When a prompt leaves open a decision that would change the result, the agent pauses, shows Waiting for answers, and holds its run until a person answers or skips, without spending credits. A run can be stopped at any point.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/make-changes-with-agent-runnersread 2026-09-22

Annual independent audits cover AICPA SOC 2 Type 2, ISO 27001, ISO 27018, PCI DSS v4.0 and HIPAA, with the SOC 2 Type 2 report available to Enterprise customers in Netlify's Trust Center. The controls customers use sit beside it. Team roles (Owner, Developer, Publisher, Internal Builder, Git Contributor, Reviewer) decide who may run agents and who may publish, SSO and SCIM come on Enterprise, and there is a team audit log. Agent runs happen in isolated environments that never touch secrets, with secret scanning on deploys.

Sourcenetlify.com/securityread 2026-09-22

Every agent run leaves a log in the Agent Runs tab with its status and credit usage per task, the files it changed and its Deploy Preview, and the Agent Runners page states that every change is tracked, reversible and auditable, with who prompted what and when. A team audit log, separate from runtime logs, records team and project changes with retention set by plan; function logs cover the customer's own agent code, and Enterprise log drains export logs to outside tools. How much step detail a run log holds is not described.

Sourcedocs.netlify.com/manage/accounts-and-billing/team-management/team-audit-logread 2026-09-22

Within a run, work carries forward: an agent run is a series of tasks, each follow up prompt continues the same run with its earlier work in place, and a new project that cannot be finished in one run ends with a reviewable portion and a plan for the remainder. Project context that a team adds applies to every later run and is read by the agent as context.

No memory that the agent writes and keeps across runs is documented, and no scope, lifetime or controls to review, edit or delete memories are described. Async Workloads keep step results for retries, but the platform applies those rather than an agent reading them.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/make-changes-with-agent-runnersread 2026-09-22

Customers can choose where the AI runtime's compute runs, though Netlify itself is a managed cloud platform with no self hosted option. The functions that run AI Gateway calls and Async Workloads deploy to US East (Ohio) by default; a menu in project settings moves them to one of several AWS regions, and EU regions in Paris and Milan are available through support, with compliance named as a reason to choose. Where Agent Runners sandboxes and AI Gateway traffic run is not stated.

Sourcedocs.netlify.com/build/functions/optional-configurationread 2026-09-22

Starter prompts for new projects come with Agent Runners, Netlify's packaged way to run coding agents on a live project, along with a prompt examples page for common jobs such as fixing broken links, redirects, content updates and new pages. Netlify Skills and context files teach outside coding agents the platform. These are working assets for one job, changing a web project, not a browsable catalog of production ready agents, templates or role specific agents for a buyer to adopt.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/overviewread 2026-09-22

Events and schedules wake the customer's agent code without a person. Async Workloads, which Netlify documents for AI work beside the AI Gateway and MCP server, start on events and can schedule or sleep between steps, and Scheduled Functions run AI Gateway code on a cron schedule. Agent Runners themselves are started by a person from the dashboard or a phone, with deploy notifications to Slack, webhooks or email.

Sourcedocs.netlify.com/build/async-workloads/overviewread 2026-09-22

Users pick both the agent and the model. Agent Runners run Claude Code, OpenAI Codex, Google Gemini or OpenCode, and each user can set the model and reasoning effort every agent uses, with OpenCode offering open models such as Kimi, DeepSeek and GLM through OpenRouter, routed only to providers with a zero data retention policy. The AI Gateway gives the customer's own code OpenAI, Anthropic, Google and open models through their official SDKs with no provider keys, and switching provider needs no rewrite. The model setting is a personal preference per user, not a team policy.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/overviewread 2026-09-22

From outside, a documented API (OAuth2) and CLI drive the platform, and an MCP server, offered as a hosted endpoint or an npm package, lets coding agents create projects, deploy sites and manage environment variables. Netlify also publishes Netlify Skills and AI context files for coding agents, and an entry point at netlify.ai for agents that deploy on their own.

Sourcedocs.netlify.com/llms.txtread 2026-09-22

Validation comes from running the agent's work: every Build mode run produces a Deploy Preview on infrastructure that matches production, with the changed files listed, and a failed build or deploy is surfaced before anything ships. That checks the work in a real environment. No way to test agent workflows against fixtures or datasets before production, score output quality over time, or compare agents or models is documented. The same preview holds the change until a person publishes it.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/make-changes-with-agent-runnersread 2026-09-22

Browser and desktop control is not documented. Agent Runners run coding agents in isolated environments that edit a project's files and build it, and a person checks the result in a Deploy Preview; no agent is shown operating a browser or desktop, clicking or typing in an interface. No browser or computer use environment is offered to the customer's own agents either, and a code sandbox that edits files does not operate web interfaces where APIs are absent.

Sourcedocs.netlify.com/build/build-with-ai/agent-runners/overviewread 2026-09-22