Hermes Wiki
AIDigest/2026/07/15/2026-07-15-1200-aws-agent-governance-stack-loom-deep-dive

Source: awslabs/loom on GitHub — repo created 2026-07-07 Also referenced: AWS Open Source Blog — "Building secure AI agents at scale: Introducing Loom for AWS" — 2026-07-08; aws-samples/sample-pluggable-agentic-ai-framework

Summary

Two recent AWS agent repos turn out to be two layers of the same stack rather than competitors. sample-pluggable-agentic-ai-framework is a notebook workshop that teaches a 5-layer reference architecture (data/knowledge → inference → orchestration → security → observability) on Bedrock AgentCore + Strands Agents. awslabs/loom is a deployable FastAPI + React control plane that operationalizes that same architecture — adding a UI, Cognito auth/RBAC-ABAC, an agent registry with approval workflows, human-in-the-loop gates, and OTel cost/latency tracking on top of the same AgentCore Runtime + Strands SDK combination.

Key Takeaways

  • Same substrate, different altitude. Both repos are built on Bedrock AgentCore Runtime + Strands Agents SDK. The workshop repo's L3 (Registry → Gateway → specialist agents) and L4 (Gateway interceptor for PII masking) map almost directly onto Loom's Agent Registry governance and MCP-server access control — Loom is the productized version of the pattern the workshop teaches by hand.
  • Loom's differentiators: on-behalf-of (OBO) token exchange (RFC 8693) so an agent acts with the calling user's scoped permissions rather than a shared service credential; four distinct HITL enforcement points (agentic loop hooks, tool context interrupts, MCP elicitation, harness inline functions); per-invocation token/cost tracking via Bedrock's CountTokens API; a staged local→hybrid(RDS)→full(ECS Fargate) deployment model.
  • Maturity check: Loom is ~1 week old (created 2026-07-07, ~105 stars at last check), Apache-2.0, explicitly "as-is" with a standard AWS sample disclaimer — worth trialing locally (Phase 1, SQLite, no AWS deploy needed) before treating it as a production dependency.
  • The circulating summary article oversold one detail — the "~40 minutes to stand up" framing is blog marketing copy, not a claim the repo itself makes; everything else in that summary (RBAC/ABAC, mandatory tagging, config-over-codegen, HITL-by-design) checked out against the actual README and blog.
  • Alternatives if AWS lock-in is unwanted: LangGraph Platform / LangSmith for observability+deployment; Backstage + custom plugins or Port/Cortex for the catalog/governance angle; a bare MCP registry + OPA (Open Policy Agent) for RBAC/ABAC without adopting the full Loom stack.

Relevance to My Work

Two pieces map onto the Network Automation/Observability domain in my professional profile (unresolved): (1) Loom's Gateway + OBO-token pattern is a ready-made template for exposing internal NetOps APIs to an agent with per-user scoped access instead of a shared over-privileged role — directly applicable to any Aegis/NetOps-adjacent agent tooling; (2) its OTel-based per-invocation cost/latency tracking is the same observability instinct already applied to non-agent systems, just scoped to LLM calls. No existing note in LocalzDocs/, CoursesDocs/, or BlogPosts/ covers agent governance/registry patterns yet, so this is a fresh thread rather than a connection to prior writing.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)

Ask about this article
Paste into a fresh claude.ai / Claude Code chat pointed at this repo (covered by your subscription, no API cost), then type your question — the exchange gets appended to this article's Discussion section and pushed.
Hermes Wiki