# Five-surface routing — Harness, Strands, Chat Agents, Flows, Automate

Use this matrix during architecture review. It is finer than the two-column [AgentCore vs Quick Suite](../agentcore-vs-amazon-q/decision-matrix.md) matrix.

> **Reproduce this** — Copy the five questions into your RFC. Pair with [`tco-scenarios.csv`](tco-scenarios.csv) and [`skills-support-matrix.md`](skills-support-matrix.md).

## Five questions (answer in order)

| # | Question | If yes → | If no → |
| --- | --- | --- | --- |
| 1 | Are users **employees** with SSO who need permission-aware answers over SharePoint / Confluence / Salesforce? | Quick **Chat Agents** | Continue to 2 |
| 2 | Is the agent **customer-facing** or embedded in **your product API**? | AgentCore **Harness** (start) or **Strands** (if topology) | Continue |
| 3 | Do you already maintain **AgentSkills.io `SKILL.md` packs** (AWS Agent Toolkit, Git, S3)? | **Harness** (`awsSkills` / Git / S3) or **Strands** (`AgentSkills` plugin) — Quick cannot consume them | Quick instructions are enough for employee chat |
| 4 | Do you need a **fixed graph**, hop caps, Swarm, or Agents-as-Tools? | **Strands on Runtime** — export Harness; do not rebuild IAM / Memory / Gateway | Stay on **Harness** |
| 5 | Must the workflow run **without a user at the keyboard**, long-running, org-owned? | Quick **Automate** (Enterprise plan only) **or** AgentCore if it is a product/ops API with Git-versioned tools | Quick **Flows** for user-run short tasks |

## Surface matrix

| Scenario | Chat Agents | Flows | Automate | Harness | Strands on Runtime | Why |
| --- | :---: | :---: | :---: | :---: | :---: | --- |
| Internal HR / IT policy Q&A | ✓ | | | | | ACL inheritance; admin console |
| Recurring “draft → review → Jira” for a team | | ✓ | | | | User-run; chat-agent step is single-turn |
| Invoice / loan / vendor process, unattended | | | ✓ | | | Enterprise-only; HITL in Automate |
| Customer storefront chat / refunds | | | | | ✓ | Product API, Gateway Policy, branding |
| First ops assistant that must follow AWS runbooks | | | | ✓ | | Official `awsSkills` glob |
| Multi-agent supervisor (sales / orders / support) | | | | | ✓ | Graph / Agents-as-Tools / hop caps |
| Employee research + BI in one workspace | ✓ | | | | | Quick Research + Quick Sight — not this pack’s SKU math |
| High-volume anonymous public FAQ | | | | ✓ | | Seat model misfits external volume |
| Team already on LangGraph they will not rewrite | | | | | keep | Runtime host; not a greenfield default |

**Legend:** ✓ = primary recommendation. Blank = do not start there.

## When NOT to choose each surface

| Surface | Skip when |
| --- | --- |
| **Chat Agents** | External users, white-label UX, or you need `SKILL.md` progressive disclosure |
| **Flows** | You need multi-turn inside one step, unattended org-owned runs, or Git-versioned tools |
| **Automate** | You are on Professional ($20) — Automate is Enterprise ($40); or the “automation” is a customer API |
| **Harness** | You already need hop caps / Graph this quarter; or the job is SharePoint Q&A for 500 employees with no engineers |
| **Strands** | A single-domain Harness still expresses the loop — do not pay container + evals tax early |

## Related

- [Skills support matrix](skills-support-matrix.md)
- [TCO scenarios CSV](tco-scenarios.csv)
- [Monday checklist](monday-checklist.md)