Manhattan Associates

Agent Foundry's Tools building block lets builders select, define and invoke ActivePlatform APIs, external APIs, foundational agents and external agents so agents interact with backend systems of record; the FAQ states tools can be APIs from any external system as well as Manhattan solution APIs that access and manipulate data or control processes; and approved service definitions encapsulate the connection, authentication and request structure for each backend system. A2A and MCP interoperability serves the other direction, letting outside systems call the agents.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Agent Foundry's Workflows block implements multi-step workflows with defined tasks and conditional logic guiding execution, the ActiveAgents page documents stateful workflows and composable agent orchestration in which multiple agents combine skills and outputs as a team, the FAQ names an orchestration engine within ActivePlatform, and builders select which agents work together on a task.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

The ActiveAgents FAQ states the reasoning engine incorporates retrieval-augmented generation and grounding, enriching the LLM with enterprise-specific data and verifying outputs against ActivePlatform data, and the agents live inside the platform's persistent system of record rather than assembling context per request from an external lake. How the retrieval index is scoped and refreshed is not described.

Sourcemanh.com/solutions/manhattan-active-platform/ai-agents-for-supply-chain-productivityread 2026-09-07

Human approval is built into the workflow: Agent Foundry's Guardrails building block implements procedural guardrails through conditional workflow logic to enforce business rules and require human approvals where necessary, the ActiveAgents page names instructional, procedural and data/API guardrails plus content sanitization as the control set, and a March 2026 article documents the agent escalating to a planner with options as a designed outcome.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Manhattan's Compliance page states that semi annual third party audits covering a twelve month period certify services against SOC 1, SOC 2 and SOC 3 and SSAE 18 / ISAE 3402 Type II, with ActivePlatform, where the agents run, and ActiveSCALE named in scope, plus annual third party penetration testing, SOX audits and PCI attestations held at the gated trust center.

On access, Agent Foundry documents an authorization scope that admins govern over the agent surface, safeguards configured by administrators and builders, API guardrails restricting builders to predefined services for authorized systems, and agents that invoke only approved service definitions carrying the connection and authentication for each backend system. That gives the customer granular control over who in its organization can make an agent do what.

No SSO, SAML, SCIM or platform wide RBAC is documented publicly; the public Security page lists control programs and the Access Control policy by title only, and the trust center at trustcenter.manh.com is gated.

SourceManhattan Associates, manh.com/trust-center/compliance and /security, Agent Foundry pageread 2026-09-07

The ActiveAgents page documents execution tracing and structured logging of agent activities with response-time and resource-consumption monitoring, and the FAQ names logging every agent decision with its underlying logic explained; a March 2026 article shows the shape, a worker reallocation logged with the utilization rates, predicted demand and productivity trends behind it. With the trace and the retained reasoning together, a reviewer can reconstruct why an agent acted, not merely what it did.

Sourcemanh.com/solutions/manhattan-active-platform/ai-agents-for-supply-chain-productivityread 2026-09-07

State lives inside a run. Agent Foundry's Data and Content building block lets builders create, modify and remove variables and parameters within the agent's memory to manage state and context, and the ActiveAgents page describes stateful workflows that carry outputs forward to later decisions. No memory layer with a stated scope (per user, team or workflow) and a stated lifetime is documented, and whether memory can be deleted is not published. The ActivePlatform data the agents are grounded on is the application's system of record rather than agent memory.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Manhattan's public trust pages (landing, Security and Compliance) and privacy policy name no hosting region, region list, customer environment option (VPC, on premises, customer tenant) or way to select one. The platform is described throughout as cloud native, evergreen, multi tenant SaaS, and the privacy policy says only that personal information may be shared with business units around the world, so a buyer with a different residency requirement has nothing to choose. The gated trust center at trustcenter.manh.com holds the security questionnaires and DPA.

SourceManhattan Associates, manh.com/trust-center/security, /compliance and /privacy-policyread 2026-09-07

The ActiveAgents page carries a browsable catalog of named prebuilt agents (Labor, Wave, Store and CSR agents on the page, with a See all Agents link to the ready-to-deploy agents catalog), the commercial-availability release names Wave Coordinator, Labor and Maven customer-service agents, the Momentum 2025 release names Intelligent Store Manager, Labor Optimizer, Wave Inventory Research, Contextual Data Assistant and Virtual Configuration Consultant, and Agent Foundry ships a template library of helper agents and reusable components.

Each agent does a distinct job a customer can deploy without the others. They are included with every ActivePlatform solution per the FAQ and the Q2 FY2026 earnings call.

Sourcemanh.com/solutions/manhattan-active-platform/ai-agents-for-supply-chain-productivityread 2026-09-07

Chat, continuous monitoring and events all start agent work. The CSR agent and Manhattan Active Maven answer inquiries and take order actions in chat; the Dock Agent, built by Eaton, monitors workflow and expedites tasks per the commercial availability release; and a March 2026 article documents an agent that predicts delays from traffic and weather signals, then cancels or rebooks appointments on its own, triggers trip replanning or escalates before an exception becomes an emergency. No schedule, webhook or inbound queue surface is separately documented.

Sourcemanh.com/solutions/manhattan-active-platform/ai-agents-for-supply-chain-productivityread 2026-09-07

Customers and admins get no control over which model runs. The ActiveAgents page says the reasoning engine uses leading LLMs, plural and unnamed, and Agent Foundry lets builders incorporate LLM calls at any workflow stage through prompt templates with no model selector documented. The Google Gemini Enterprise and Agentspace integrations publish agents to Google's marketplace, which is extensibility rather than model choice.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Outside callers can reach the agents several ways. Agent Foundry documents agents exposed as callable tools to be discovered and orchestrated by Google Gemini Enterprise, published to an agent marketplace, registered as callable foundational agents for other agents, and callable with standard APIs, alongside A2A and MCP support.

The product pages name ActivePlatform's own microservices APIs as the surface agents and customers call, and a Developer Hub at developer.manh.com and a Developer Tools page are linked from the primary navigation. The A2A support names no agent card, endpoint or auth scheme publicly.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Measurement happens before and after deployment. The ActiveAgents FAQ says customers compare process outcomes before and after deploying an agent, a March 2026 article says the platform tracks decision speed, error reduction and operational outcomes before and after deployment so a buyer can see whether the agent earned its keep, and Agent Foundry manages the lifecycle from creation and versioning to retirement.

That is a live metric against a baseline with no control group, so it measures change over time rather than the agent itself, and versioning is lifecycle management rather than evaluation. No harness, scored test cases, release gate or controlled experiment after deployment is documented, and the vendor's language about continuously creating, testing and deploying agents has no testing surface behind it. The approval gate before execution checks each action rather than a change to an agent.

Sourcemanh.com/solutions/manhattan-active-platform/build-ai-agents-agent-foundryread 2026-09-07

Agents act through microservices APIs, approved service definitions and external APIs inside and around Manhattan Active, and no browser, desktop or remote computer control is documented on the ActiveAgents, Agent Foundry or ActivePlatform pages.

SourceManhattan Associates, manh.com platform pagesread 2026-09-07