Scaled Cognition

Typed actions against the customer's own API schemas connect agents to the customer's real systems. Under "APIs & typed actions", the platform page says to "connect real API schemas so agents execute with verified contracts", and the Agent Builder lets teams "test against simulated APIs... then deploy to real systems".

Meet APT describes validating "every parameter against your system schemas before executing any function", and APT "confirms successful system execution" before telling a customer an action is complete, with actions on account balances, order statuses and transfers. "Typed APIs specify exactly what each agent can read, write, or modify", and custom tools are the customer's own API schemas, not only prebuilt connectors.

Named native connectors and credential rotation or revocation are not documented. The Genesys Cloud partnership is a distribution channel for the model, not the agent's integration.

Sourcescaledcognition.com/product/explore-platform, /product/meet-apt-1; readread 2026-09-15

Agent behavior compiles into deterministic specifications that mix fixed conditions with agent steps, with versioned deployments. Behavior is authored as structured instructions in natural language, a low-code builder or the Python SDK, "all compiled into deterministic specifications instead of prompts", with deterministic conditions and function conditions that "specify... preconditions that must be satisfied before APT is permitted to predict an action".

That mixes deterministic nodes with autonomous agent steps and gates each action on conditions, which is routing and branching at the step. Changes are versioned and debuggable: "controlled evolution" with "versioned deployments, evaluation gates, and validation", plus auditable execution. Business, product and engineering teams author on one deterministic core.

Explicit retry and fallback paths are not described.

Sourcescaledcognition.com/product/explore-platform, /product/meet-apt-1; readread 2026-09-15

Grounding comes from live data read from the customer's systems at run time, with no maintained index. Under "Zero guessing", Meet APT says APT "retrieves facts, doesn't generate them" and "grounds responses in real-time information from your systems: account balances, order statuses, policy details". The platform page has agents retrieve verified data, and the SDK enforces data provenance. Knowledge is read at run time, not trained in.

Reads through typed APIs at the moment of a conversation are context assembled per run, however many sources they draw on, and the platform states "stateless processing, no storage, no training, no retention", so no persistent index over the buyer's content is described. Which sources are indexed and refreshed, and citations shown to a customer, are not documented.

Sourcescaledcognition.com/product/meet-apt-1, /product/explore-platform; readread 2026-09-15

Policy is enforced through the architecture rather than through human sign off. The design promises "strict policy adherence", with "business rules enforced through architectural constraints, not prompts", so agents "physically cannot execute unauthorized actions"; function conditions set preconditions before any action; Agent Defender blocks jailbreaks, malicious prompts and policy violations in real time before actions execute; and "APT confirms execution, not intent". These are structured guardrails and policy constraints at strength.

The buyer cannot decide which actions need sign-off, who signs off, or what happens when nobody does. No human approval step, sign-off routing or pause control is documented; the design replaces sign-off with enforced preconditions. Generative virtual customers that test escalations are a testing tool, and the vendor's team verifying agents before go live is an engagement service.

Sourcescaledcognition.com/product/meet-apt-1, /product/explore-platform; readread 2026-09-15

A SOC 2 Type 2 attestation and per agent access control are both stated. On compliance, the Go Live page's Security & Compliance tile reads "SOC 2 Type 2 compliant, controls are designed into the platform environment", in present tense and naming a report type, and the vendor runs a Vanta trust center at trust.scaledcognition.com.

On access, under "Deterministic access control", "typed APIs specify exactly what each agent can read, write, or modify", and function conditions gate which actions are permitted, which is least-privilege tool access. Platform tools process interactions ephemerally, with "no storage, no training, no retention".

No SSO, SAML, SCIM or user roles are documented for people using the builder, and the SOC 2 report's auditor and date are not stated on the public pages.

Sourcescaledcognition.com/product/go-live, /product/explore-platform, trust.scaledcognition.com (metadata only); readread 2026-09-15

Every action, blocked decision and policy check is logged, and live agent behavior is monitored. Under "Auditable execution", the platform page says "log every action, blocked decision, and policy check, providing complete traceability for compliance, debugging, and oversight", and adds "every interaction within the platform is logged and auditable".

The Go Live page states "every agent action is traceable, with clear visibility into what was allowed, blocked, and why". Behavior monitoring in live customer conversations catches "issues that aggregate metrics miss", and Agent Monitor tracks "drift, policy violations, and unexpected actions". That is run-level visibility, including the steps taken and the tools called.

SIEM export and retention by plan are not documented. Because stateless processing with zero data retention is stated for the platform tools, how long audit logs themselves persist is unstated. Trend & Topic Analysis reads conversation content for business strategy, not agent behavior.

Sourcescaledcognition.com/product/explore-platform, /product/go-live; readread 2026-09-15

No state carries between conversations, by stated design. Under "Flexible deployment", agents run "with stateless execution, zero data retention", and under Compliance by Design the platform promises "stateless processing, no storage, no training, no retention", adding that "platform tools process interactions ephemerally". No memory beyond the conversation in progress is documented. AgentTwin learning from historical transcripts builds an agent specification, which is authoring, not memory the agent keeps.

Sourcescaledcognition.com/product/explore-platform; readread 2026-09-15

Deployment can be hosted, in the customer's own VPC or on premises. Under "Flexible deployment", the platform page says "run agents hosted, in VPC, or on-prem, all with stateless execution, zero data retention", and again under Compliance by Design: "deploy on your infrastructure (hosted, VPC, or on-prem) to maintain complete data sovereignty". The Go Live page adds "deploy in your VPC or on-prem, with sensitive data staying inside your environment". Data placement is settled by running in the customer's environment.

Features by mode, who patches in each mode, and hosted regions are not documented.

Sourcescaledcognition.com/product/explore-platform, /product/go-live; readread 2026-09-15

No prebuilt agents, templates or catalog are documented. AgentTwin "learns from what your best agents already do, and ships it as a production-ready agent, straight from your transcripts", which is a generator that builds a custom agent from the customer's own data; nothing is prebuilt or selected. The Vibe Builder, Studio and SDK are authoring surfaces. The site map lists industry pages (telecommunication, travel and hospitality, banking and finance, retail, BPO and CX service providers, healthcare).

Sourcescaledcognition.com/product/explore-platform, sitemap; readread 2026-09-15

Live customer conversations reach the agents, but the channels they arrive on belong to a partner's platform, not Scaled Cognition's own. The vendor's blog announces that Genesys "has selected Scaled Cognition's specialized customer experience large action models to power virtual agents on the Genesys Cloud platform". The inbound queue and channels in that deployment belong to Genesys Cloud, a third party's contact center platform.

Scaled Cognition's own product pages name no channel (voice, chat, email, messaging), inbound queue, webhook, event or schedule that starts an agent, even though contact center deployment implies inbound traffic.

Sourcescaledcognition.com/blog/announcing-our-partnership-with-genesys (16 October 2025), /product/explore-platform; readread 2026-09-15

Every agent runs on APT, Scaled Cognition's own model, with no provider choice. The vendor is a model lab: APT is "the only CX model built on super-reliable intelligence", compared on the Meet APT page against general LLMs, and every authoring route is "powered by the same deterministic core: APT". No provider list, routing policy, or bring-your-own-model or key path is documented. Genesys Cloud offering APT to its customers is distribution of this model, not model choice inside Scaled Cognition's platform.

Sourcescaledcognition.com/product/meet-apt-1, /product/explore-platform; readread 2026-09-15

A Python SDK is named for the platform, but no documentation for it is published. The APT SDK is described as "engineering-grade control for technical teams. Script, extend, and integrate agents in Python, enforce data provenance, and test behavior", and appears on the Explore Platform, Meet APT and Go Live pages alongside the Vibe Builder and Studio. That is a real developer surface.

No developer, API reference or SDK page is in the site map, and the API host at api.scaledcognition.com sits behind an access gate. No MCP surface is mentioned.

Sourcescaledcognition.com/product/explore-platform, /product/meet-apt-1, sitemap, api.scaledcognition.com probe; readread 2026-09-15

A full evaluation stack produces readable, comparable results. For simulation, GenAPI simulation emulates the customer's infrastructure "without connecting to live systems", generative virtual customers simulate realistic and adversarial behavior, and a test case generator builds scenarios from standard workflows to edge cases.

For results, "agent scorecards measure correctness, policy adherence, and reliability across repeated runs" and track them over time, and evaluation validates agents against thousands of simulated scenarios "to catch performance regressions before deployment". Changes go through "controlled evolution", which uses "versioned deployments, evaluation gates, and validation". Fixtures and datasets before production, quality gates, scoring over time and a post-deployment loop through Agent Monitor are all documented.

The expert led evaluations on the Go Live page are an engagement service rather than part of the product.

Sourcescaledcognition.com/product/explore-platform, /product/meet-apt-1; readread 2026-09-15

Scaled Cognition is built on the opposite premise from computer use, so there is no browser or desktop control. Agents act through "APIs & typed actions" with "verified contracts", tested against GenAPI simulations of backend APIs, and "APT validates every parameter against your system schemas". With no screen in the loop, changes to UI elements cannot break it.

Sourcescaledcognition.com/product/explore-platform, /product/meet-apt-1; readread 2026-09-15