AWS Step Functions workflows become AI agent tools via AgentCore Gateway

AWS now lets developers register existing Step Functions state machines, both Standard and Express, as tools in Bedrock AgentCore Gateway. The design splits responsibilities cleanly: the LLM decides whether to call a workflow and with what input, and Step Functions then runs every step exactly as defined, orchestrating any of more than 220 AWS services.

The pattern addresses a core enterprise worry about agents: non-determinism in business-critical processes. By putting audited, versioned workflows behind a tool interface, teams keep their existing orchestration logic and compliance posture while letting agents choose when to invoke it. AWS published a MoEngage case study alongside the launch: the company turned roughly 400 private REST APIs across 7 regions into agent tools using OpenAPI specs, Cedar policies and VPC Lattice egress, avoiding per-service MCP servers, public endpoints and context bloat from hundreds of tools.

This is AWS's governance-first agent strategy in action, and it complements the AgentCore CLI 1.0 release earlier in the week. Microsoft is pushing a parallel story with its Copilot harness in Foundry Agent Service. The fight is increasingly over the control plane around models rather than the models themselves.

What to watch: pricing as agent-triggered workflow executions scale, and whether customers actually prefer deterministic tools over letting agents improvise.