Quick verdict
Most “AI agent development cost” pages sell a blob: $8k–$400k tiers with no roadmap when you add agent #3. Cipher quotes a production job as three line items after a scope call:
- Stamp — the production-ready platform we already run. Identity, MFA, access control, audit logging, isolation. Paid once; every agent reuses it.
- Per production agent — marginal cost to harden one proven or clearly scoped agent into production.
- Per connector / integration — each external system (Microsoft Graph, HubSpot, and similar) as its own line so integration risk stays visible.
The public floor is the job, not the formula: a typical 2-agent production job starts from AUD $40,000. Bands live on /pricing/. Best for: Australian and Singapore teams with Claude or ChatGPT workflows that already work — or clearly scoped agents. Honest limit: we do not sell ISO 27001 / SOC 2 certification, and we do not fixed-price unbounded prompt-churn.
Last updated: August 2026. Cipher Projects is an Australian AI engineering studio that builds production agents and the platforms they run on.
Why blob quotes fail buyers and delivery teams
A single fixed number feels simple until the second connector and the third agent show up. Then someone eats the overrun: either the client pays surprise change-orders, or the studio under-delivers and delivery quality collapses.
Transparent line items are a trust mechanism, not a marketing slogan. You can check the commercial architecture before you sign. We can staff and schedule what we actually promised. Delivery stays sustainable because scope is named.
The three-line model (indicative AUD)
| Line | What it is | Indicative floor | Why it exists |
|---|---|---|---|
| Typical 2-agent job | Two production agents and two connectors on the stamp | From AUD $40,000 | The public floor — what /pricing/ sells |
| Stamp | Shared rails every agent reuses | Quoted after a scope call | Paid once so scale does not rebuild identity and runtime |
| Per production agent | Harden one agent (design, tools, evals, UAT, runbook) | From AUD $12,000 (typical ~$12–15k) | Scales predictably with agent count |
| Per connector | One external system wrapper + auth + tests | From AUD $3,500 | Stops burying integration risk inside a blob |
USD and SGD conversions appear in the questions on the pricing page and in /pricing.md. Greenfield work with no stamp is a higher project in the same bucket. Product companies reselling to downstream clients quote higher. Per-agent stays in the same band either way.
What the stamp actually includes
The stamp is the difference between a Claude chat that “works for the team” and a system operators can audit:
- Cloud landing posture that does not block ISO 27001 / SOC 2 controls
- Unique users, MFA required, access control, short-lived sessions
- Authenticated UI and an application layer in front of the agents
- Shared memory, observability, and a first connector wrapper
Architecture supports and does not block ISO 27001 / SOC 2. Certification is a third-party audit and excluded. Runtime model bills are cloud spend — separate from the build quote. See Bedrock pricing if the job lands on AWS.
What moves price up or down
| Direction | Drivers |
|---|---|
| Up | New connectors the agent depends on; human-in-the-loop approval chains; external customers on the login path; heavier compliance review; agents not yet proven in Claude/ChatGPT |
| Down | Agents already proven with real usage; shared tools/memory across a family; read-only agents (no writes to external systems) |
Unique insight from quoting work: the fastest way to blow a “cheap agent” quote is burying two SaaS connectors and an undefined third agent inside one fixed blob. Separating connectors is how both sides stay honest when HubSpot OAuth or Graph consent takes longer than the prompt work.
Claude workflow vs production agent (pricing lens)
| Claude / ChatGPT workflow | Production agent on your platform | |
|---|---|---|
| Identity | Vendor account / project share | Your SSO/MFA, per-user tool scoping |
| Isolation | Chat session | Per-session runtime isolation + audit trail |
| Connectors | Manual paste or thin MCP | Gateway wrappers, OAuth, least privilege |
| Cost control | Seat + usage surprise | Model routing, timeouts, eval sampling budgets |
| Human approval | Person in the chat | Explicit escalation paths in the workflow |
| Commercial path | Subscription | Foundation once → agents N → connectors later |
Full cutover narrative: Claude workflow vs production agent.
Worked shape (anonymised)
Two agents already proven as Claude workflows, Microsoft 365 as first connector, one additional SaaS connector, internal staff only, SOC 2 / ISO 27001 aspirations (infra that does not block — not the cert itself):
- Stamp on the production-ready platform we already run
- Two per-agent lines
- Two connector lines
- Typical phase from AUD $40,000 · roughly 8–12 calendar weeks with two senior AI engineers. Greenfield (no stamp) quotes higher.
That is a shape, not a promise. Intake changes the point estimates; the line-item structure does not.
What is not included
- Certification: we prepare controls and evidence paths; assessors certify your organisation.
- Unbounded discovery: “we will invent the agents as we go” belongs in advisory or a managed seat, not a fixed platform quote.
- Cloud runtime bills: model tokens, runtime meters, vector/storage — operational spend.
- Managed monthly seats: ongoing capacity is a different product on /pricing/.
FAQ
Is this cheaper than a managed AI engineer? Different job. Seats are recurring capacity from AUD $9,500/month. The agent job is a fixed-scope build. Many clients ship the first agents, then keep a seat for iteration.
Can we start with one agent and add more later? Yes — that is the point of the stamp. Later connectors slot in without rewriting identity and access control.
Do you only build on one cloud? AWS is a common path for AU/SG teams that need isolation and audit language. The public quote is the job, not a runtime brand.
Who is this for? Founders and operators who already feel the Claude workflow working and need production identity, connectors, and a quote they can defend to a board — not a vendor blob.
Conclusion
Price the job the way you will operate it: stamp once, agents as you prove them, connectors as named systems. Public floor: typical 2-agent job from AUD $40,000 on /pricing/.
Related: Claude workflow vs production agent · AgentCore pricing · Bedrock pricing · Applied AI engineering