L4 is multi-agent. Specialized agents (planner, researcher, writer, reviewer, executor) coordinate via shared state, handoffs, and a control plane. Runs durably across hours or days, with checkpointing, parallelism, retries, audit. The "agent fleet" becomes a real org chart. "AI ops" becomes a function.
Technically: a graph of agents (LangGraph, CrewAI, AutoGen/AG2) or a managed platform (Bedrock AgentCore, Vertex Agent Builder, Mosaic Agent Bricks, Cortex Agents), with shared memory, evals as the heartbeat, observability, and edge-gated escalation.
Not human-in-the-loop as default. HITL is the exception, not the rule: if the evals can't tell whether the output is right, fixing the evals is the priority, not stapling a reviewer to the loop.
L4 is also where lock-in becomes a real exit consideration. Bedrock, Vertex, Agentforce, and Mosaic Agent Bricks create durable platform dependencies. Edge Scale is being built specifically to keep the orchestration control plane portable across these runtimes, so a buyer doesn't inherit a vendor.
We separate the orchestration framework (the harness, chosen for control and inspectability) from the runtime (where it executes and how it's governed, in-tenant for regulated clients). Keeping them decoupled lets us change the harness without re-platforming, and swap the platform without rewriting the harness.
Chosen for control & inspectability.
Chosen for governance & data proximity.
Bedrock AgentCore
Mosaic AI
Azure AI Foundry| Function | In practice | Signal |
|---|---|---|
| Support | Agent-run tier 1 and tier 2; humans escalate only on edge cases | Outcome-priced |
| Sales | An SDR agent pool: research → send → reply classification → meeting booked | Fleet scale |
| Finance | AP agent: invoice intake → matching → approval routing → exception escalation | Approve exceptions only |
| HR | Recruiting agent: inbound screening → scheduling → first-round scorecard | Recruiter sees finalists |
| Engineering | Multi-Devin or Claude Agent SDK fleet working a backlog you can't hire against | Parallel SWE agents |
| Category | BOD Pick | Why | Lock-in |
|---|---|---|---|
| Multi-agent framework | LGLangGraph + LangSmith | BOD L4 default. Checkpointing, time-travel debug, graph viz. 110K+ stars, 35% of Fortune 500. | Low |
| Multi-agent (alt) | CRCrewAI | Role/crew metaphor lands with stakeholders. Insight-backed. Managed Enterprise. | Low |
| MS-aligned | AGAutoGen / AG2 | Strong conversation patterns. Microsoft-adjacent procurement. | Low |
| Anthropic-native | CLClaude Agent SDK | The cleanest path when Claude is primary. | Low |
| AWS-managed | Bedrock AgentCore | Fully managed. Memory + KBs + action groups. AWS-only. | High (AWS) |
| GCP-managed | GCVertex AI Agent Builder | ADK + Agent Studio + 200+ models including Claude and Gemini. | High (GCP) |
| Databricks-managed | Mosaic AI + Agent Bricks | Auto-tunes against benchmarks. Presupposes Lakehouse. | High (Dbrx) |
| Snowflake-managed |  Cortex Agents | Snow-native. Easiest path if Snow is system of record. | High (Snow) |
| MS-managed | Copilot Studio + Azure AI Foundry | Most-adopted enterprise platform (38.6% per JetBrains 2026 survey). | Very high |
| Vertical CX | DEDecagon | $4.5B val. Notion, Duolingo, Substack logos. Proven multi-agent CX. | Medium-high |
| Vertical SWE | DVCognition Multi-Devin | Parallel autonomous SWE. Pricey; needs supervision. | Medium |
| Evals heartbeat | BTBraintrust + OTel | Quality as engineering. OpenTelemetry as substrate. | Low |
Edge Scale is BOD's proprietary Agent Ops control plane, a product-in-build, with capabilities phasing through 2026. Orchestration-agnostic by design: it runs on top of LangGraph, CrewAI, Bedrock, Vertex, Cortex, or Mosaic, and stays cloud-portable.
Governance, audit, cost attribution, RBAC, SSO, connector catalog, deployment blueprints, and a managed tier targeting 99.9% SLA. "The architecture is the asset."
Sierra (outcome-priced), Decagon (multi-agent enterprise), or LangGraph-built with Edge Scale on top. Human escalation only at policy edges.
A fleet of research → outreach → reply classification → meeting agents, with per-agent budgets and a shared eval harness. Pipeline scales without headcount.
Invoice intake → matching → approval routing → exception escalation. Finance approves exceptions, not invoices.
Inbound screening → scheduling → first-round scorecard. The recruiter takes only shortlisted candidates into final rounds.
Parallel autonomous SWE agents working the backlog the team can't hire against. Supervised by senior engineers; not autonomous in the L5 sense.
Onboarding agent runs sign-up → first value. CS agent watches usage, flags risk, drafts save motions. Pricing model starts shifting seat-based → outcome-based.
Governance, audit, cost attribution, RBAC, SSO, connector catalog, deployment blueprints: a managed tier targeting 99.9% SLA, orchestration-agnostic by design. Cloud-portable across Databricks Mosaic AI, Snowflake Cortex, and AWS Bedrock AgentCore.
Talk to BOD