Selector Foundry lets NetOps teams author AI agents on their own telemetry

Selector introduced Selector Foundry, a build-and-run environment inside its network platform so operators can write, test, version, and promote their own AI agents where the network data already lives, without copying that data to an outside agent service.

SOFTWARE desk — a vendor-written network agent cannot know which nightly link flap is normal for this team. Foundry turns the operators’ runbooks into software that starts from a named cause, instead of copying the network’s data to someone else’s agent cloud.

What Foundry is, in the page’s words. It is a development and runtime environment. Development is where a team writes the agent. Runtime is where that agent runs. It lets network operations teams build, test, version, and govern their own AI agents inside the Selector platform. Network operations, often shortened to NetOps, is the team that keeps the network running. An agent, here, is software that takes steps, not only a chat reply. Version means each saved edition of that agent. Govern means the team controls which edition is allowed to run. Agents execute where the telemetry is already collected, so no operational data is copied to an external agent service. Operational data is the record of how the network is behaving. An external agent service would be someone else’s product that needs a copy of that record. These lines are the company’s. This desk did not install Foundry.

Where an agent starts. Selector’s platform resolves signals from network, cloud, infrastructure, and application sources into a single causal picture with a named probable cause, rather than a ranked list of possibilities. A signal, here, is one of those incoming records. A causal picture is one account of what led to the problem. A named probable cause is the one reason the platform names, not a menu of maybes. Until now, the page says, the work operators did after that correlation stayed manual. They assembled evidence. They opened and chased the ticket. They pulled the right teams onto a bridge. They decided on remediation. A ticket is the record in the help system. A bridge is the call where those teams meet. Remediation is the fix. Each of those steps, the page says, increases time to resolution, which is how long the problem stays open. Foundry agents take on that sequence. Because they consume Selector’s correlated output rather than raw logs, they begin from the evidence instead of working the cause out again. A log is the raw line-by-line record. Each response from the orchestrator agent contains a recommended action and the evidence behind it. An orchestrator, here, is the agent that returns that package. Operators still decide whether to act. The page says the response moves faster without the risks of full automation. Full automation would mean the software acts without that human decision. Faster is the company’s word. This desk did not time a ticket.

Why the company says the operators have to write the agents. No two networks are the same, and Selector’s position is that this makes operator authorship a requirement, not an optional feature. Two enterprises running identical hardware still differ in topology, vendor mix, naming conventions, maintenance windows, escalation paths, and what each team considers normal. Topology is the shape of the network: which box connects to which. A vendor mix is which companies made the gear. A naming convention is how the team labels those boxes. A maintenance window is the planned hour when a change is allowed. An escalation path is who gets called when the first person cannot fix it. The example on the page: a link that flaps nightly is an incident in one environment and expected behavior in another. A flap is a link that drops and comes back. An incident is a problem the team treats as an outage. A vendor-written agent has no way to know the difference, because that knowledge lives in the operations team’s runbooks and in the judgment of the engineers who have run the network for years. A runbook is the written procedure the team already follows. Foundry makes that knowledge executable. The people who already know which conditions matter are the people who write the agents that act on them. File that example as the company’s. This desk did not watch a nightly flap.

How the page says a team ships one. The senior operator is the author, and the output is treated as software. Agents are defined, committed to the customer’s own Git repository, and reviewed through the team’s existing pull-request workflow. A commit is a saved version. A pull request is a proposed change a teammate reviews before it is accepted. Before promotion, an agent is replayed against the customer’s own historical event data. The replay produces a pass or a fail against the recorded outcome of each past incident. Pass means the agent’s result matched what the record says happened. Fail means it did not. A failed promotion is rolled back in a single step. Rollback means the promotion is undone. Agents can start from an event, from a schedule, from a REST API call, or from an operator’s request. A REST API is a web address another system can call to start the agent. They run under the RBAC and identity controls Selector already enforces. RBAC is role-based access control: the rules for which person, or which agent, may do which job. Identity is who is signed in. These lines are the company’s. This desk did not open a repository or replay an incident.

Six agents ship with the runtime, and what the page actually names. Foundry ships with six generally available agents, so a team is not starting from an empty runtime. Generally available means the company says those six are on for customers, not held in a closed trial. The page describes the work. They investigate incidents from the correlated evidence. They open and update records in IT service management systems, which the page shortens to ITSM, and they attach the root-cause context. ITSM is the ticket system. Root cause is the reason the platform named. They generate decision-ready views and stakeholder reports. A stakeholder report is a write-up for people who need the outcome, not the raw logs. They operate across cloud estates, covering asset health, connectivity, and spend. A cloud estate is the set of rented computers and services a company runs. Asset health is whether those pieces are up. Connectivity is whether they can reach each other. Spend is what they cost. The page says six. It does not print six product names. Do not invent them. All six arrive in the same repository as agents the customer writes. A team can run them unmodified or fork them to match its own runbooks. Unmodified means as shipped. Fork means copy one and change it. File the list as the company’s.

The quote, as the company’s, not as a measured result. Nitin Kumar, co-founder and chief executive, said other organizations are building agent platforms and asking customers to send their network data to those platforms. He said Selector put the agents where the network already is, so a team builds them the way any other software gets built: they design it, review it, test it, and promote it when it is ready. He said the agents are the customer’s, in the customer’s repository, on the customer’s data. File the name, the title, and those sentences as the post’s. A quote is not a timed resolution, and it is not a customer list.

The about box, and where it stops. Selector calls itself an AI-powered observability and network intelligence platform that unifies data, correlation, and automation across domains. A domain, here, is one kind of system: the network, the cloud, the machines, or the applications. The box says the platform combines large language models, knowledge graphs, and causal reasoning so teams can detect, diagnose, and resolve issues faster. A large language model is software that reads and writes text. A knowledge graph is a map of how the records connect. Causal reasoning is the named-cause step already described. The box says leading telecommunications providers, cloud service providers, and global enterprises rely on Selector to reduce MTTR, prevent outages, and accelerate transformation. MTTR is mean time to repair: how long a broken service stays broken, on average. Leading, and those customer categories, are the company’s. The page does not name a customer. Do not add one. Selector is backed by Ansa Capital, Atlantic Bridge Ventures, AT&T Ventures, AVP, Bell Ventures, Comcast Ventures, Hyperlink Ventures, Two Bear Capital, Sinewave Ventures, and Singtel Innov8. That list is the company’s. This desk did not check who holds the shares. The page does not print a price. Do not add one.

What the shipped frame shows, and what it does not prove. The product image with this filing is a Selector workflow panel. The labels on that panel, in order, are Investigate, Validate, Propose fix, Test impact, and Prepare rollback. Those words are on the image. The newsroom text does not print that five-step list. Do not treat the panel as a second announcement, and do not treat a mark on the image as a result this desk measured.

Plain English for the rest of the card: Foundry = the build-and-run environment announced on this page. NetOps = the team that keeps the network running. agent = software that takes steps, not only a chat reply. telemetry = the measurements and events the network already sends. operational data = the record of how the network is behaving. external agent service = someone else’s product that would need a copy of that record. named probable cause = the one reason the platform names. orchestrator = the agent that returns a recommended action plus the evidence. flap = a link that drops and comes back. runbook = the written procedure the team already follows. Git = the version-control store. pull request = a proposed change a teammate reviews. replay = run the agent on past incidents and compare with what the record says happened. rollback = undo a promotion in one step. REST API = a web call that can start the agent. RBAC = who is allowed to do what. ITSM = the ticket system. generally available = the company says the six agents are on for customers. fork = copy an agent and change it. six = the page’s count. The page does not print six product names. MTTR = how long a broken service stays broken, on average. This filing is the 22 Sep announcement.

PRIMARY here: Selector’s 22 Sep 2026 newsroom post, “Selector Foundry Brings AI Agents to Network Operations, Enabling Organizations to Build, Version, and Run on Their Own Data,” datelined Santa Clara — Tier A PRIMARY, the company’s own record. The page did not print an hour. The Foundry name, the build-test-version-govern line, the no-copy line, the named probable cause, the manual sequence the agents take on, the orchestrator’s recommended action plus evidence, the operator’s decision, the flap example, the Git repository and pull-request review, the replay pass or fail, the one-step rollback, the four triggers, the RBAC line, the six generally available agents and the jobs the page describes without six product names, the unmodified-or-fork line, the Kumar title and quote, the about-box methods, the MTTR line, the unnamed customer categories, and the investor list are the post’s. The five step labels on the product frame are on the image, not a second list in the newsroom text. NOT claimed: that this desk installed Foundry, ranked a leader, timed a ticket, watched a flap, opened a repository, replayed an incident, named six products the page left unnamed, named a customer, found a price, treated the image as a measured result, a stock tip, or investment advice. Distinct from the already-filed digitalocean-managed-agents, teradata-tera, and salt-security-aidr.