ADR-0003: Anthropic Managed Agents — Build vs. Adopt Evaluation
Status: Accepted
Date: 2026-04-10
Decision: No scope change for v1.0. Selectively adopt MCP proxy and code execution sandbox in v1.1+.
Context
Anthropic released Managed Agents (beta) — a hosted service virtualizing the agent loop into Session (state), Harness (orchestration), and Sandbox (execution). Beta API: /v1/agents, /v1/sessions, /v1/environments.
Decision
Continue building as planned. Our multi-tenant platform layer (Context Engine, portal, billing, workflows, compliance) is differentiated and unaffected. Managed Agents is headless single-tenant infrastructure; UpsQuad is a multi-tenant SaaS platform built on top of such primitives.
What Managed Agents does NOT provide (our differentiators)
- Multi-tenant isolation (org > pillar > team > agent, RLS)
- Role-based access control (L1-L5 clearance)
- Hierarchical agent teams
- Workflow engine (approval chains, SLA)
- Billing/pricing (dual billing, feature gates)
- Client Portal (chat UI, dashboards, onboarding)
- Compliance/audit (per-tenant audit trail, GDPR)
- Guardrail engine (injection detection, chain of trust)
- Cost optimization (savings metrics, budget alerts)
Selective Adoption Plan
| Area | Decision | When |
|---|---|---|
| Context Engine | Keep ours | — |
| MCP tool integration | Adopt their proxy | v1.1+ |
| Code execution sandbox | Adopt their environments API | v1.1+ |
| Streaming responses | Evaluate proxying their stream | v1.1+ |
| Agent runtime loop | Evaluate if multi-tenant added | v2 |
Risks
- Beta API may change
- No pricing announced — cost structure matters for margin
- Vendor lock-in — mitigate with abstraction interfaces