Is the model locked in, or can we swap providers?

We route through a model gateway, so the agent calls a provider as a string and you can swap between Claude, GPT, Gemini, or a future model with a config change. Your prompts, tools, and evals stay portable — we avoid SDK patterns that would tie you to one foundation model.

What stops the agent from doing something harmful or expensive?

Scoped, audited tool permissions, prompt-injection defense, cost caps, rate limits, and human-in-the-loop escalation for ambiguous cases — all built in, not bolted on. You can see traces and eval scores from day one of production.

How do we know it's actually working in production?

Three numbers, reviewed weekly: task success against the golden set, average cost per resolved task, and human-escalation rate. If any drifts past the agreed threshold, we pause and root-cause before rolling further.

What is the Pilot-to-Production program?

A fixed-scope 90-day enterprise program: one agent into production behind your IAM with an eval harness, observability, and an ISO-42001-aligned governance package. Your team learns the operating model on the way through, so the next agent costs less.

Build · 1–10 Weeks

AI Agent Build

An engineered, tested AI agent running in production.

Startup

Simple from $10k

Company

RAG-workflow $18–25k

Enterprise

Multi-agent $30–50k+ · Pilot-to-Production from $165k

Timeline

Simple 1–2 wks · RAG-workflow 3–5 wks · Multi-agent 6–10 wks · Pilot-to-Production 90 days

What You Get

An agent with its prompt, tools, and integrations built and tested
A golden test set and an offline eval harness you can re-run
Prompt-injection defense and scoped, audited tool permissions
Cost caps, rate limits, and runaway protection
A production runbook and an incident SLA

Packages

Simple

from $10,000

A single-purpose agent — one job, a few tools, in production.

Prompt + tools + integration
Golden test set + offline evals
Cost caps + production deploy

RAG Workflow

$18,000–$25,000

An agent grounded in your knowledge with retrieval and workflow logic.

RAG corpus + retrieval evals
Multi-step workflow + tool-use
Injection defense + audit logging

Multi-Agent / Enterprise

$30,000–$50,000+ · Pilot-to-Production from $165,000

Multiple coordinated agents, or a governed 90-day pilot-to-production program with ISO-42001-aligned documentation.

Multi-agent orchestration + observability
Behind your IAM/SSO with audit logging
Governance package (AI inventory, risk, monitoring) on the enterprise program

In Scope

Discovery

Weeks 1–2

Use-case selection and impact sizing
Data inventory, sensitivity classification, access mapping
Eval-harness design — golden set, synthetic edge cases, red-team prompts
Architecture: model choice, RAG vs tool-use, deployment surface

Build

Weeks 3–8

Agent on AWS Strands, model-independent via a gateway
MCP integration layer with audit logging and scoped permissions
RAG indexing with retrieval evals, where applicable
Observability: traces, evals, cost telemetry, anomaly alerts

Harden & Ship

Weeks 9–10

Eval pass against golden and red-team sets
Cost ceilings, rate limits, human-in-the-loop escalation
Production deploy behind your IAM, runbook, knowledge transfer

How It Works

A scoping call, then a Blueprint Sprint locks scope, tier, and a fixed price.

A named senior pod builds against a fixed timeline — no hourly billing.

Production deploy behind your auth, full IP, optional Run & Care after.

What “Done” Means

Your agent runs in production behind your auth, with eval scores you can see.

Why Codenovai

We build agents behind evals, guardrails, and audit logs — the control system that keeps them stable — on the same framework and patterns we run in our own products.

Fixed price

Shipped in weeks

Code + IP in your repo, day one

No lock-in

FAQ

Is the model locked in, or can we swap providers?: We route through a model gateway, so the agent calls a provider as a string and you can swap between Claude, GPT, Gemini, or a future model with a config change. Your prompts, tools, and evals stay portable — we avoid SDK patterns that would tie you to one foundation model.
What stops the agent from doing something harmful or expensive?: Scoped, audited tool permissions, prompt-injection defense, cost caps, rate limits, and human-in-the-loop escalation for ambiguous cases — all built in, not bolted on. You can see traces and eval scores from day one of production.
How do we know it's actually working in production?: Three numbers, reviewed weekly: task success against the golden set, average cost per resolved task, and human-escalation rate. If any drifts past the agreed threshold, we pause and root-cause before rolling further.
What is the Pilot-to-Production program?: A fixed-scope 90-day enterprise program: one agent into production behind your IAM with an eval harness, observability, and an ISO-42001-aligned governance package. Your team learns the operating model on the way through, so the next agent costs less.

Ready to start?