Multi-Agent Orchestration Engine
Build complex agent pipelines with session management, warm pooling, and sandboxed execution. Chain agents together into production workflows that scale.
What is Forge Foundry?
Forge Foundry is the orchestration engine for multi-agent workflows. While ForgeBot lets you build individual agents, Foundry lets you wire them into pipelines -- sequences of agents that hand off work to each other, branch based on results, and fan out across multiple specialists in parallel.
Think of it as a workflow engine where every step is an AI agent. Each agent gets its own isolated session with scoped memory, tools, and cost budgets. Foundry handles the plumbing: session lifecycle, credential forwarding, warm pool management, and full trace observability for every pipeline run.
Core Infrastructure
Everything you need to run agent pipelines reliably at scale, without managing the underlying infrastructure yourself.
Pipeline Architecture
Define multi-step agent pipelines as declarative configurations. Each stage in the pipeline runs an agent with specified inputs, outputs, and failure handling. Pipelines support branching, looping, and parallel fan-out.
Session Management
Every Foundry session gets an isolated execution context with its own memory scope, tool permissions, and cost budget. Sessions persist across pipeline stages so agents share context without leaking state between tenants.
Warm Pool Management
Pre-provisioned agent instances sit ready to execute. When a pipeline stage fires, Foundry pulls from the warm pool instead of cold-starting a new agent. Pool sizes auto-scale based on traffic patterns and time-of-day demand.
Auth Bridging
Foundry forwards the calling user's subscription tier and API credentials into each agent session. Agents inherit permission scopes without separate authentication, and cost is metered back to the originating account.
Sandboxed Execution
Each pipeline stage runs inside resource-limited boundaries. CPU, memory, network, and execution time are all capped. If an agent exceeds its budget, the stage is terminated gracefully with a partial result returned to the pipeline.
Observability Built In
Every pipeline execution produces a full Langfuse trace. You can see which agents ran, how long each stage took, what each agent decided, and how much it cost -- all from a single trace ID.
Orchestration Patterns
Foundry supports multiple execution patterns. Combine them within a single pipeline to model any workflow, from simple chains to complex decision graphs.
Sequential Chains
Run agents one after another. The output of Stage A becomes the input of Stage B. Ideal for research-then-write, analyze-then-recommend, or extract-then-transform workflows.
Conditional Branching
Route pipeline execution based on intermediate results. If an analysis agent flags a high-risk item, branch to a deep-dive agent. Otherwise, skip ahead to summarization.
Parallel Fan-Out
Split a task across multiple agents running simultaneously. Fan out a research query to three specialists, then fan in their results for synthesis. Cuts latency for complex multi-perspective tasks.
Human-in-the-Loop Gates
Insert approval checkpoints into any pipeline. The pipeline pauses, notifies a human reviewer, and resumes only after explicit approval. Critical for high-stakes workflows like financial approvals or code deployments.
Foundry Marketplace
Pre-built pipeline templates for common workflows. Install, customize, and deploy -- or publish your own pipelines for others to use.
Browse & Install
Find pre-built pipelines in the Foundry Marketplace. Install a customer onboarding pipeline, a code review workflow, or a content production chain with a single click.
Customize & Fork
Every marketplace pipeline is a starting point. Fork it into your workspace, swap out agents, adjust prompts, add stages, or change branching logic to fit your exact process.
Publish & Earn
Built a pipeline that works well? Publish it to the Foundry Marketplace. Set a price or offer it free. Earn revenue every time another team installs and runs your pipeline.
Build Your First Pipeline
Read the documentation to learn how Foundry pipelines work, then create your first multi-agent workflow.