Switchboard Cuts Automation Build Time in Half with Claude on Amazon Bedrock

Building reliable multi-agent automations typically requires hand-coding agent orchestration, tool integrations, and failure recovery putting automation out of reach for teams without deep engineering resources and slowing down even experienced teams. GLIB built Switchboard so users can assemble automations visually, or describe the desired outcome in natural language, without hand-coding orchestration, tool integration, or durability logic themselves.

Switchboard is built around five blocks: Data Sources (the platform’s connectable database), Agents (autonomous components configured with intake, a system prompt, tools, runtime limits, and output format), Workflows (canvases connecting nodes, agents, and conditions across a catalog of 35+ pre-built nodes spanning control flow, integrations, document processing, extraction, and data operations), Build with AI (a natural-language assistant that generates or modifies workflows, agents, prompts, and tools from a described requirement), and Cases (a testing area to run and inspect automations step by step before publishing). Automations run as durable Temporal workflows, agents discover and call tools via the Model Context Protocol (MCP), and agents can delegate to sub-agents via Agent-to-Agent (A2A) delegation. Claude Sonnet 4.5 powers Build with AI as a fixed default, while individual agents within a workflow can be configured to run on Claude Sonnet 4.5, Claude Haiku, or another model depending on the task; Amazon Bedrock is the default inference path, with direct provider APIs available as a configurable alternative.

At the core of Switchboard is Claude. Claude Sonnet 4.5 powers Build with AI which is the system that turns a described requirement into agents and workflows as a fixed model, so every automation is built with the same high-capability model regardless of deployment. Individual agents running inside a published automation are configurable per task, defaulting to Claude Sonnet 4.5 with the option to route lighter-weight work to Claude Haiku. Both are accessed through Amazon Bedrock by default, with direct provider APIs available as a configurable alternative.

Model: Claude Sonnet 4.5 on Amazon Bedrock is a fixed default for Build with AI. Agent runtime within a published automation defaults to Claude Sonnet 4.5 and can be configured to Claude Haiku or another model per task. Direct Anthropic API access is available as an alternative inference path to Bedrock.

What Switchboard does today

• Data Sources – a connectable database of Entity or Collection tables, with typed, constrained columns and a defined primary key, usable directly by any agent or workflow.
• Agent building – each agent is configured with typed intake (fields and file slots), a structured system prompt (role, scope, procedure, output protocol, and guardrails), runtime limits (model, max iterations, temperature, max tokens, LLM and tool timeouts), access to the platform’s full node catalog including other agents as sub-agents and a defined output format (text, JSON, or custom).
• Workflow building – automations are assembled visually from a catalog of 35+ pre-built nodes across nine categories including the ability to call external APIs or invoke another workflow as a sub-step.
• Build with AI – a natural-language assistant that generates or modifies workflows, agents, prompts, and tools from a described requirement, usable inside an active workflow or as a standalone chat with context from existing workflows and agents.
• Testing and publishing – automations are run and inspected step by step in Cases, with errors flagged for correction, before being validated, saved, and published as a new, shareable, exportable version.

Measured results attributable to Claude on Bedrock

• 2x faster to build a comparable multi-agent automation, per internal engineering comparison.
• 40+ automations published and running in production.
• 120+ active builders using the platform without engineering support.
• Median time from idea to published automation: 3 business days, down from an estimated 6 business days building the equivalent by hand.

Responsible AI

Every agent’s system prompt includes an explicit guardrails component with rules and limits the agent must always respect alongside a defined scope statement separating what is and isn’t within the agent’s remit. Execution is further bounded at the runtime level: a configurable maximum-iterations limit caps how many reasoning/tool loops an agent may run, and separate LLM and tool timeouts bound how long any single call may take. A human review step can be added directly into a workflow before a consequential action executes, an audit trail is maintained for platform activity, and tenant isolation is enforced between different users’ automations.

Switchboard gives GLIB’s users a visual, no-code path to building multi-agent automations: data sources, agents, and workflow logic are assembled on a canvas, or generated directly from a natural-language description through Build with AI, rather than hand-coded. Claude Sonnet 4.5 powers Build with AI as a fixed default, so every automation is built with the same high-capability model regardless of deployment. The agents that then run inside a published automation are configurable per task defaulting to Claude Sonnet 4.5, with the option to route lighter-weight, high-volume steps to Claude Haiku instead. Amazon Bedrock is the default access route to Claude, with direct provider APIs available as a configurable alternative.

Underneath the visual builder, automations run as durable Temporal workflows, so a multi-step agent process can recover from failure rather than restart from scratch. Agents discover and call tools including other agents, used as sub-agents through the Model Context Protocol, and delegate work to those sub-agents through Agent-to-Agent communication. Every agent’s system prompt carries its own guardrails and scope, execution is bounded by configurable iteration and timeout limits, and a human review step can be inserted into any workflow before a consequential action executes. Platform activity is captured in an audit trail, and automations belonging to different users are isolated from one another.

Since going live, GLIB teams have published 40+ automations on the platform, with roughly 120 active builders most without an engineering background assembling and iterating on workflows directly. Building a comparable automation from scratch, including environment setup, dependency resolution, and pipeline work, takes roughly twice as long as building the same automation on Switchboard, with the trade-off of less flexibility than a fully custom build.