What happened
Levanto Labs announced Sage2 on November 2, 2026, describing it as extraction and streaming decisions for software that needs to act on human language. The announcement frames the problem as an application receiving a customer's changing request: the system needs details for an offer, a decision about what to do, and a way to keep both current when the customer changes their mind. Sage2 provides two separate API operations that can be used together.
The extraction API lets a developer define the fields an application needs and their types. In Levanto's example, a customer says they want a midsize SUV for up to $450 a month. The request asks for vehicle, monthly limit and loan term; Sage returns the vehicle and budget as structured values, while the term is null because it was not specified. Adding “over 60 months” produces P60M, an ISO 8601 duration. Supported types are text, numbers, money, dates, times and durations. The announcement says each value includes the text it came from and a probability, and is linked to its position in the input. An application can inspect the source, check results and request missing information before calling a tool.
The streaming API connects from a server to a WebSocket endpoint. A developer opens a session, defines questions, then appends only new text as it arrives; Sage maintains accumulated context and returns answers as the conversation changes. Levanto's example asks whether a customer has explicitly accepted the current offer and asked for a quote. An offer is not acceptance, and an objection leaves the answer at no; accepting a revised offer changes it to yes. Sessions support up to 32 questions and 16,384 tokens of context and questions. The page says replay timing is illustrative.
The company presents the APIs as composable: extract requirements, search eligible offers, keep checking whether the customer wants to proceed, then let application code prepare a quote when the answer changes. Sage supplies fields and decisions; the language model responds; the application validates an offer and executes the next step. Levanto says the same pattern can apply to support requests, document processing and agent tool selection. Sage2 also supports reasoning, web grounding and vision in beta. Extraction and streaming decisions are available today to holders of an existing Sage API key, with a playground and documentation linked from the announcement. The page gives no price.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This is an instance of model answers becoming task interfaces: instead of treating a response as free-form prose, Sage2 returns typed values and bounded decisions that application code can inspect and use. The example is a vendor's product design, not evidence that this interface is becoming a universal standard.
Why it matters
Conversational workflows often fail at the boundary between understanding a request and taking a consequential action. A transcript may contain a budget, missing details and a change of mind, but an application still needs explicit values and a reliable trigger for what happens next. Sage2 makes that boundary visible: developers define fields and questions, inspect extracted evidence, and keep execution in application code. That separation can make it easier to test whether the system captured the right value and whether the customer actually agreed before preparing a quote or invoking a tool. It does not remove the need for review; it gives teams a more structured surface on which to build it.
Limits and uncertainty
Levanto's announcement describes examples and API limits, but does not report extraction accuracy, decision error rates, latency, uptime or independent evaluations. It does not state pricing, rate limits, retention terms or how probabilities should be calibrated for a particular application. The demo's replay timing is explicitly illustrative. The company says reasoning, web grounding and vision are in beta, but the announcement does not detail their scope. Teams should test their own inputs and keep consequential validation in their code.
Skill to build now
Take one existing conversational workflow and write down the fields it needs, including how each field should behave when absent. Then list the yes-or-no decisions that can change as new messages arrive. Use representative transcripts to test extraction, missing-value handling and decision changes before connecting any result to a tool or customer-facing action. That mirrors Levanto's own recommendation to start with one workflow and evaluate results before automating.
Reader takeaway
Sage2 is not a replacement for application logic: it is an API layer for turning dialogue into typed values and updating explicit decisions. If your product acts on changing human requests, the practical idea is to separate extraction, decision checks and execution—and verify each boundary before automating the next step.