# Harness vs Runtime vs DIY LangGraph/CrewAI

Decision matrix for net-new production agents on AWS (August 2026).

| Dimension | AgentCore Harness | AgentCore Runtime + Strands | DIY LangGraph / CrewAI on Runtime or ECS |
| --------- | ----------------- | --------------------------- | ---------------------------------------- |
| Ownership | Config: model, prompt, tools, memory | You own agent code; AWS owns microVM host | You own loop, retries, packing, deploy |
| Multi-agent | Single managed loop; export when topology outgrows config | First-class Agents-as-Tools, Graph, Swarm, Workflow + A2A | Bring your own supervisor / graph |
| Time to first agent | Minutes (CLI `agentcore create` or CreateHarness) | Hours–days (container/CodeZip + IAM) | Days–weeks (loop + state + observability) |
| Graduation path | Export to Strands code (same Memory/Gateway/Identity) | Already code | Port tools into Gateway; keep framework |
| When to pick | Greenfield single-domain agents; ≤ ~10 tools; config is enough | Explicit topologies, hop caps, A2A, custom orchestration | Existing LangGraph/CrewAI investment you will not rewrite |
| When NOT to pick | Compliance DAG with hard step order; multi-team A2A | Thin FAQ bot Harness already covers | Greenfield where Harness + export would remove plumbing |

## Opinionated default

1. Start **Harness** for the first production path.
2. **Export to Strands** when you need multi-agent topology or custom loop control.
3. Keep **DIY LangGraph/CrewAI** only if the graph already exists and the cost of rewrite exceeds Gateway/Memory rewiring.

## Cost note

Model tokens usually dominate. Cap Browser / Code Interpreter on every turn — platform compute can 3× support-bot bills when those tools stay hot. Model platform spend on the [AgentCore pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/).