AgentCore Harness + Strands vs Amazon Quick: Chat Agents, Flows, and Automate (August 2026)

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic and Amazon Q Business are in maintenance for new customers after July 30, 2026. Net-new custom agents should use Bedrock AgentCore. Net-new workforce AI should use Amazon Quick Suite. Full matrix: lifecycle roundup.

On August 14, 2026, the procurement question is no longer “AgentCore or Quick.” AgentCore Harness has been generally available since June 17, 2026 (What’s New, Harness skills). Strands Agents 1.0 already ships Graph, Swarm, Workflow, and an official AgentSkills plugin. Amazon Quick splits the employee side into Chat Agents, Flows, and Automate (Flows, Automate vs Flows). Pick the surface, not the brand name.

This post is the five-lane comparison. It is not a rewrite of the July 2026 platform TCO guide (workforce assistant vs custom agent infra) or the August Harness ship map. Those still stand. What they do not answer: whether Agent Skills are official on Quick, whether a Flow can hold a conversation, and what 500 employee seats cost on the August Quick price list.

First-party signals we reuse (not new client invoices) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Support-style AgentCore at 50K sessions/mo ~$791/mo platform + model — July TCO worksheet. August Quick list-price remodel of the same 500-employee silhouette: Professional ~$10,250/mo — tco-scenarios.csv.

Reproduce this — Copy examples/architecture-blog-2026/agentcore-harness-strands-vs-quick/ into the RFC. Fill seat counts in tco-scenarios.csv. Route with decision-matrix.md. Confirm live rates on Quick pricing and the AgentCore calculator.

Opinionated take: Quick Chat Agents and Flows win time-to-value and ACL connectors for employees. You give up AgentSkills.io portability and customer-facing APIs. Harness wins the first production product or ops agent that must load AWS procedures. Strands on Runtime wins the day config cannot express hop caps. Automate is Enterprise-only unattended process work — not a Strands substitute.

History: how five names landed on one RFC

If you only remember one line from the Quick Suite history post: Quick is the workforce workspace. AgentCore is how you run agents you own. The rest of this page is which surface inside those two brands.

Five surfaces (stop collapsing the names)

Harness and Strands load Agent Skills. Chat, Flows, and Automate do not. Diagram is a label aid, not an AWS architecture icon set.

Quick Flows vs Automate (AWS’s own split): Flows are personal/team, user-run, shorter tasks. Automate is centralized, can run without a user, long-running, extra control-flow and custom code — and it is not on Professional (choosing Flows vs Automate).

Harness vs Strands: treat Harness as the paved road for the first production agent. Treat Runtime + Strands as the paved road once topology, hop caps, or A2A show up. Export Harness; do not rebuild IAM, Memory, or Gateway. That graduation is the ship map.

Agent Skills: official on Harness and Strands, not on Quick

This is the lock-in tell the July platform guide did not cover.

Agent Skills are bundles of markdown and scripts in the AgentSkills.io shape: a SKILL.md with YAML frontmatter, optional scripts/ and references/. Metadata is injected up front (~100 tokens); the full body loads on demand. That is progressive disclosure.

Full table with source URLs: skills-support-matrix.md. The AWS Agent Toolkit is the catalog Harness awsSkills is built to consume.

Assumes AWS CLI v2, a Harness execution role, and a supported region. Empty awsSkills object enables the curated bundle; glob paths if you want a subset.

# Enable AWS-curated skills on a harness (control plane)
aws bedrock-agentcore-control create-harness \
  --harness-name ops-assistant \
  --execution-role-arn "${ROLE_ARN}" \
  --skills '[{"awsSkills":{"paths":["core-skills/*","specialized-skills/operations-skills/*"]}}]'

# Per-invoke overlay — invoke-time skill with the same name wins
# boto3 with bedrock-agentcore client; pin the SDK your account documents
response = client.invoke_harness(
    harnessArn=HARNESS_ARN,
    runtimeSessionId=SESSION_ID,  # >= 33 characters
    skills=[{"awsSkills": {"paths": ["core-skills/aws-cdk"]}}],
    messages=[{"role": "user", "content": [{"text": "Draft a CDK stack for a Lambda function."}]}],
)

Do not port SKILL.md packs into Quick instructions. That drop loses globs, Git pins, and on-demand loading. If the team already maintains skills, the surface is Harness or Strands. If they do not, and the users are employees asking SharePoint, Quick instructions are the right artifact — just do not call them skills.

Cost: July worksheet vs August list price

One worked example (the standard allows one silhouette for math): 500 employees on Quick vs one 50K-session/month product agent on AgentCore.

Sources: Quick pricing fetched 14 August 2026 (Professional $20, Enterprise $40, $250/account infrastructure, included agent hours Pro 2+2 / Ent 4+4, overage $3/agent-hour and $6/research-hour, index 25 GB / 50 GB pooled per user then $5/GB-month). AgentCore row C is reused from the July worksheet, not re-benchmarked. Plug your seats into tco-scenarios.csv.

Honest gap: we do not have a published Quick Automate client invoice. The Enterprise row is list price on the same 500-employee shape, labeled as such.

What changed vs July: Lite $3 and index $/hour are the wrong quote for a net-new AWS-account Quick buy in August. Professional is the floor for custom Chat Agents and Flows on that page. Automate is the reason you pay $40. Agent hours are the silent overage — Research, Flows, Automate, desktop, and custom apps meter to the second after the included pool.

AgentCore bill shape is unchanged: Harness rides Runtime-shaped vCPU/GB-hour + Memory events + optional Gateway/Browser/Code Interpreter + Bedrock tokens. Skills fetch is not a separate SKU. Browser left on still dominates — see the 12-component pricing post and the calculator.

Direction of the July result still holds and the gap widened for “give every employee a seat” vs “meter a product agent.” Right-size seats (row H) before you declare Quick too expensive.

Feature updates that change the choice

Each bullet is who breaks, not a feature list.

• Harness GA + Skills (17 Jun 2026). Teams still copying runbooks into a 4k-token system prompt. Turn on awsSkills or a Git pin instead.
• Strands 1.0 Graph / Swarm / A2A. Harness config cannot express hop caps. Export; do not wait for a checkbox.
• Quick Chat Agents / Flows / Automate split. Automate is Enterprise-gated. A Professional quote that “includes automations” is a Flows quote unless you said $40.
• Flows chat-agent step is single-turn (AI response steps). Multi-turn interviews in a Flow fail UAT. Stay in Chat or move to Harness.
• Agents Classic / Q Business maintenance (30 Jul 2026). Net-new Classic is the wrong default. Existing tenants keep running.

Three implementation shapes (choose, don’t tour)

1. Customer storefront — Strands on Runtime

The eCommerce AgentCore sample is a supervisor plus four specialists, eight Gateway tools, Cedar on refunds and cancels. That is a product API with Policy-gated writes. Quick Chat Agents cannot be the shopper storefront: no white-label session contract, no Cedar on Gateway, seat math does not fit anonymous volume.

Choose Strands (Runtime). Do not start on Harness if you already know you need Agents-as-Tools and hop caps. Do not start on Quick.

2. Internal HR / IT helpdesk — Chat Agents + Flows

AWS’s public employee onboarding chat agent pattern is the right shape: Spaces for the handbook, action connectors for tickets, share with a group. Our July FAQ still holds: if the goal is permission-aware answers over ServiceNow, Confluence, and SharePoint in 30–60 days and you have no agent engineering team, AgentCore is over-engineering.

Add Flows for the repeating “open the ticket, draft the reply, post to Slack” path. Stay on Professional unless the process must run unattended.

Choose Chat Agents + Flows. Move a write-back that Flows cannot express to Harness — not the whole helpdesk.

3. Ops assistant that must follow AWS runbooks — Harness + awsSkills

If the agent’s job is “how do we handle an ECS service stuck in DRAINING” against your runbooks and AWS-recommended procedures, that is the AWS Agent Toolkit plus Harness core-skills/* and specialized-skills/operations-skills/*. Quick instructions are not a substitute for the catalog.

From a real engagement — A B2B CRM assistant, 12 Gateway tools, ~8k turns/day. Moving tool execution server-side on AgentCore Gateway cut median tool round-trip ~180 ms → ~95 ms. That latency budget is why we will not put the same tool chain behind a Quick action connector and call it done. Detail: Gateway server-side tools.

What broke — Support-bot pilots on AgentCore. Browser stayed enabled on conversational turns that only needed Gateway tools. Platform compute landed at roughly 3× a Runtime-shaped session. Detected on the vCPU/GB-hour line, not in eval scores (the answers still looked fine). Recovery: default Browser and Code Interpreter off; enable per skill or per intent. Same warning is in the Harness ship-map FAQ.

When to choose (five questions)

Answer in order. Full RFC copy: decision-matrix.md.

1. Employees + SSO + ACL chat? → Quick Chat Agents.
2. Customer or product API? → AgentCore Harness (or Strands if topology is already known).
3. Existing SKILL.md packs? → Harness or Strands. Quick cannot consume them.
4. Graph, Swarm, hop caps? → Strands on Runtime. Export Harness; keep Gateway and Memory.
5. Must run with no user at the keyboard? → Quick Automate (Enterprise) or AgentCore if it is a product/ops API you version in Git.

Hybrid is the enterprise default: Quick for the workforce, AgentCore for the product. Shared Identity Center, separate evals, separate cost alarms.

What to Do This Week

1. Write one sentence: which surface this RFC is buying. Not “we will evaluate AgentCore and Quick.”
2. Copy monday-checklist.md into the ticket.
3. Remodel seats on August Quick list prices in tco-scenarios.csv. Do not paste the July ~$3,580 into a 2026-Q3 budget.
4. Inventory SKILL.md repos. If any exist, attach them on Harness or Strands this week — do not paste into a Quick persona.
5. If the agent is a product API, stand up one Harness in non-prod and cap Browser. If it is employee Q&A, create one custom Chat Agent against one Space.
6. Need a two-week surface-choice review (Quick employees + AgentCore product)? Book it or start from Generative AI on AWS.

What This Post Doesn’t Cover

• Kiro / Amazon Q Developer — IDE coding assistants; they compete with Copilot, not Harness. See the July guide’s Q Developer FAQ.
• Bedrock Agents Classic migration runbooks — inventory tools and sessions; not a one-day flip. Use the production AgentCore guide.
• LangGraph / CrewAI on Runtime — keep them if you already have a graph you will not rewrite this quarter. They are not the greenfield default next to Strands 1.0.
• A new Quick Automate client bill — we do not have one published. Enterprise list price is modeled, not measured.
• AWS Context — still Coming soon as of the Harness ship map; do not sequence grounding behind it.
• HIPAA/PCI control maps — both lineages can sit under a BAA; that is a separate engagement, not a SKU picker.

Need the RFC filled in with your seat count and tool inventory? Start with Amazon Quick Suite consulting or contact us.