Quick verdict
If your team already runs a working Claude Project, Custom GPT, or ChatGPT workflow, you have proven decision logic. You have not yet shipped a production agent.
- Claude / ChatGPT workflow — Best for discovering prompts, tools, and human-in-the-loop habits with staff in a vendor UI. Honest limit: identity, data boundary, connector privilege, and audit are the vendor’s product story, not yours.
- Production agent on your platform — Best for agents that must run under your SSO, write to your systems with least privilege, survive deploys, and leave an audit trail operators trust. Honest limit: you pay for foundation rails and named connectors; demos that skip those lines are not the same deliverable.
Cipher’s cutover rule: keep iterating in Claude until the workflow has real usage and a clear tool list; then harden onto a platform foundation once, price each agent and each connector as its own line. Commercial model: how we price production AI agents · bands on /pricing/.
Last updated: August 2026. Anthropic continues to ship Projects, connectors, and Claude Code routines — useful for MVP velocity. None of that replaces your AWS (or equivalent) identity and gateway story for regulated AU/SG operators.
Side-by-side: what “works in Claude” still lacks
| Capability | Claude / ChatGPT workflow | Production agent |
|---|---|---|
| Identity | Vendor seat / project share | Cognito or Entra SSO + MFA; app RBAC |
| Session isolation | Chat thread | Per-session runtime isolation (e.g. AgentCore) |
| Tool credentials | User paste, shared keys, or vendor MCP | Per-user OAuth / vaulted secrets; least privilege |
| Connectors | Ad hoc | Gateway wrappers with tests and consent flows |
| Memory | Project files / chat memory | Short-term session + long-term store you control |
| Observability | Vendor UI | CloudTrail + agent traces you can retain |
| Evals | Spot checks | Regression suite + adversarial / PII cases |
| Human approval | Person in the chat | Explicit escalation steps in the workflow |
| Cost control | Seat + surprise usage | Model routing, timeouts, sampling budgets |
| Commercial path | Subscription | Foundation once → agents N → connectors later |
When the Claude workflow is enough
Stay in the vendor UI when:
- Only internal staff use it, and a wrong answer is annoying, not costly
- Writes to source systems are still manual (human copies the output)
- You are still discovering the decision tree — tools and edge cases change weekly
- There is no compliance or customer-data boundary that requires your own audit trail yet
That is a valid MVP. Cipher often prices proven-in-Claude agents at the bottom of the per-agent band because discovery is already done.
When you need a production agent
Cut over when any of these are true:
- The agent must act (write) in CRM, email, document stores, or lending systems
- Downstream customers or partners will log in
- You need SSO, retention, or evidence language that maps to SOC 2 / ISO 27001 controls
- You want a second and third agent without rebuilding identity and UI each time
- Finance asks what agent #3 will cost — and “another blob quote” is not an answer
Unique insight from delivery quoting: teams that skip naming connectors in the SOW almost always rediscover OAuth consent, polling-only APIs, or missing webhooks mid-build. That is why connectors are separate line items on our pricing page — not because the API call is exotic, but because integration risk is where schedules slip.
Cutover checklist (Claude → production)
- Freeze the happy path — inputs, decision logic, outputs, who triggers it.
- List every system the agent touches — each one is a connector candidate.
- Mark write vs read — writes need harder evals and often human approval.
- Define the identity story — staff-only SSO vs external users.
- Agree eval cases — including adversarial and PII leakage attempts.
- Set cost controls — model tier for triage vs high-stakes output; idle timeouts.
- Price foundation once, then agents and connectors as lines — see the pricing pillar.
Where Anthropic’s newer surfaces fit
Claude Projects, Artifacts, connectors, and Claude Code routines improve MVP speed. Routines can run on Anthropic-managed (or self-hosted) infrastructure with schedules and API triggers — still inside Anthropic’s product boundary for identity and governance.
Use them to prove value. Do not confuse “runs without my laptop” with “runs under our Entra tenant with our CloudTrail and our connector consent.” For AWS-bound AU/SG teams, Cipher’s default harden path is Strands + Bedrock AgentCore behind a NestJS BFF — details in the pricing pillar and AgentCore meters.
FAQ
Can we just wrap the Claude API and call it production? You can ship a thin API wrapper quickly. You still owe identity, tool privilege, evals, observability, and connector hardening — otherwise you have a chatbot with a private URL.
Is n8n or Zapier the production layer? Great for automation glue. Weak when you need per-session agent isolation, shared agent memory, and compliance-shaped audit. Cutover guide: n8n agents vs custom frameworks.
How long does the harden phase take? Foundation plus two proven agents and a couple of connectors commonly lands in an 8–12 week phase with two senior AI engineers — scoped after intake, not promised from a blog.
Who should build it? An engineering studio that will publish the commercial architecture up front. Cipher Projects does that work for Australian and Singapore teams — Applied AI engineering.
Conclusion
Claude workflows earn the right to become agents. Production is the platform underneath: identity, isolation, connectors, evals, and a quote shaped as foundation + per agent + per connector. Keep the MVP where it is until the checklist is honest — then harden once.
Related: How we price production AI agents · Agent platform pricing · Mastra vs Strands · Durable workflow frameworks