PART 02

The Responsibility Split

Managed Agents supplies an agent harness and runtime options. It changes the engineering surface; it does not remove product, security, privacy, or operational work.

Reviewed 31 Aug 2026 · beta architecture · no universal timeline

Sources: Anthropic's overview and self-hosted shared-responsibility model. Re-check the beta docs for your deployment choice.

What Moves — and What Stays

Delivery time depends on integrations, risk, data, and evidence. Use this responsibility map instead of a universal implementation deadline.

Agent loop and session lifecycle PLATFORM HARNESS
Cloud sandbox provisioning MANAGED OPTION
Session events and server tool execution PLATFORM SERVICE
Authorization and business guardrails YOU OWN
Tenant isolation, privacy, retention, and legal basis YOU OWN
Evaluation, monitoring, incident response, and rollback YOU OWN
Platform can reduce: HARNESS & RUNTIME PLUMBING
Still required: PRODUCT & SECURITY ENGINEERING

n8n / Make.com Comparison

A fair framing. Different tools for different problems.

n8n / Make.com
A workflow-orchestration product for explicit graphs, integrations, schedules, retries, and human steps. Its available code, file, and AI capabilities depend on the deployment and nodes you configure.
Workflow Automation
Managed Agents
A configurable model-driven harness for longer, less deterministic tasks. Tools run only when enabled; permissions, network scope, application-enforced custom tools, and approvals define what it can actually do.
Agent Harness

Choose by task shape, control needs, and evidence — and combine deterministic workflows with bounded agents when useful.

Why Marketers Need This

Marketing work mixes deterministic data movement with ambiguous analysis. Start read-only, keep claims traceable, and require approval for changes to spend, targeting, publishing, or customer communication.

Ad data is messy
↓
Validate schemas and provenance; let the agent propose exceptions for review
Reporting needs repeated reconciliation
↓
Pilot read-only summaries and measure accuracy, latency, and reviewer effort
Campaign creation is manual
↓
Generate drafts in a sandbox; a human approves before any platform write
Creative testing is slow
↓
Use licensed assets, human review, and controlled experiments — never predicted lift as proof