Deep Agents Deep Dive
How Deep Agents workloads run under Faramesh governance, including delegation, MCP boundaries, and operational rollout.
Deep Agents Deep Dive
Section titled “Deep Agents Deep Dive”This page explains how to run Deep Agents workloads under Faramesh governance, where Faramesh intercepts execution, and how to roll out delegated and MCP-backed workflows safely.
Core integration model
Section titled “Core integration model”Deep Agents are governed by wrapping the runtime command with Faramesh:
faramesh run -- python -m deep_agents.mainThat runtime path is the primary enforcement boundary documented in the Faramesh core technical guide.
What Faramesh governs
Section titled “What Faramesh governs”- Pre-execution policy checks for tool and action dispatch.
- Permit, deny, and defer decision handling before the action runs.
- Centralized audit evidence for downstream review.
- Delegated and sub-agent activity when it crosses governed execution boundaries.
Source-backed technical notes
Section titled “Source-backed technical notes”The Faramesh core technical guide maps Deep Agents support to:
LangGraph dispatch + AgentMiddleware
That means the important surfaces to validate are graph dispatch and middleware-driven execution paths, not just the top-level agent loop.
Delegation guidance
Section titled “Delegation guidance”When a Deep Agents workflow delegates work to a sub-agent:
- Keep the supervisor scope explicit.
- Limit tool scope and action ceilings.
- Treat delegated actions as governed boundaries rather than free-form execution.
- Verify audit output for the delegated path, not just the parent task.
MCP interoperability
Section titled “MCP interoperability”If the workflow uses MCP tools, route them through the Faramesh MCP boundary so they do not bypass governance:
faramesh mcp wrap -- node your-mcp-server.jsThe underlying MCP interception specification covers the lower-level JSON-RPC and hardened HTTP surface.
Operational rollout checklist
Section titled “Operational rollout checklist”- Start with one small task.
- Confirm the task produces audit entries.
- Test delegated/sub-agent execution.
- Add stricter rules for risky tools.
- Re-run the same task and confirm the expected deny/defer behavior.
Validation commands
Section titled “Validation commands”Use these checks from the Faramesh core docs when validating the runtime path:
go test ./internal/adapter/mcp ./cmd/faramesh ./internal/daemongo test ./...Source references
Section titled “Source references”faramesh-core/docs/power-users/frameworks/DEEP_AGENTS_TECHNICAL_GUIDE.mdfaramesh-core/docs/guides/frameworks/DEEP_AGENTS_QUICK_GUIDE.mdfaramesh-core/docs/power-users/FEATURES_TECHNICAL_REFERENCE.mdfaramesh-core/README.md