What is Agent Blueprint?

Agent Blueprint gives AI agents one plain page: what an agent can read, what it can change, and who has to say yes. Anyone can have a coding agent draw its own blueprint. For companies, it is also a methodology and platform for designing and deploying AI agent teams, producing a living blueprint for the organization: an AI readiness assessment, a ranked use-case portfolio, agent team designs, deterministic business cases, and implementation plans. Blueprints are machine-readable, so coding agents like Claude Code, Codex, and Cursor build from them directly.

How does Agent Blueprint work?

The method runs in stages: capture the organization's business profile, assess AI readiness, rank candidate use cases by impact and feasibility, design agent teams with orchestration patterns and guardrails, build a deterministic business case, and plan a phased rollout with test gates. Humans review every stage. After deployment, agent performance is tracked against the business case and the blueprint is updated, not archived.

Is Agent Blueprint self-serve?

The personal blueprint is: paste one line into your coding agent, with no sign-up, at agentblueprint.ai/personal. Enterprise work is delivered through hands-on engagements. Client teams get platform access, API tokens, the CLI, and the MCP server as part of an engagement. To see how it would work on your organization, use the contact form at agentblueprint.ai/#contact.

Who is Agent Blueprint for?

Anyone who runs a coding agent and wants a plain picture of what it can do, and organizations that want AI agents in production with evidence behind every decision. For organizations, we work directly with your team during an engagement, and your coding agents consume the blueprint through Agent Skills exports and the MCP server.

How do I install the Agent Blueprint MCP server?

Install via npm: npx agentblueprint. The server connects a coding agent to your organization's blueprint registry, with tools for listing and downloading blueprints, business cases, and implementation plans. It uses stdio transport and works with any MCP-compatible agent. An API token issued with your engagement is required.

What open-source skills does Agent Blueprint publish?

Free, Apache 2.0 licensed Agent Skills at github.com/agent-blueprint/agent-blueprint-skills: foundation skills (agent-fundamentals, servicenow-ai-agents, agent-deployment, research-agent) and blueprint patterns for common agent teams (IT service desk triage, procurement RFx processing, customer onboarding). Install via npx skills add agent-blueprint/agent-blueprint-skills.

What makes a blueprint different from a consulting deliverable?

A typical strategy deliverable is a document that starts aging the day it lands. An Agent Blueprint is a living artifact in a registry: recommendations link to readiness evidence, the blueprint exports as files coding agents build from, and performance data flows back in after deployment. When the business changes, the blueprint changes with it.

Measured runs, 2026-11-02

What everyday AI agent jobs cost

We gave Claude Code and Codex ten realistic jobs and ran each one five times. All 100 runs got the job right. A typical job cost $0.09 to $0.35 at API prices. Cost followed how much the agent had to read, not how hard the job was.

By Chris Faulkner, Agent Blueprint. Runs on 2026-11-02 with Claude Code 2.1.307 (claude-opus-5-5) and Codex CLI 0.154.0 (gpt-6-astra). Agent Blueprint is the method NowGentic uses in its paid agent delivery work.

How to estimate your own job

Pick the closest job below and scale by how much more or less your agent has to read. Each job says how big its input was, and they're small: real emails, threads and documents are usually longer.
Multiply its typical input and output tokens by your model's current prices. Most input is cached, about 81% for Claude Code and 88% for Codex here, and cached input costs far less.
Allow for a worst case. Most jobs' most expensive run was within 1.3x of typical, but one data-cleanup run used 2.3x.
Expect a floor even for tiny jobs, about 29k input tokens for Claude Code and 60k for Codex, because the agent re-reads its own instructions and tools at every step.

Results per job

Typical (median) of five runs, with the worst run in brackets. Claude Code cost is what it reports at Anthropic API rates. Codex reports no dollars on a ChatGPT plan, so its cost is estimated from its tokens at OpenAI's standard API prices for gpt-6-astra on 2026-11-02: $10 per million input tokens, $1 cached, $50 output.

SQL questions
Answer five questions from a small SQLite shop database (150 customers, 1,200 orders).
Claude Code: $0.09 (worst $0.12). 46k input tokens (worst 59k), 1,861 output, 4 tool calls, 21 s. Passed 5 of 5.
Codex: $0.13 est. (worst $0.17). 61k input tokens (worst 61k), 673 output, 3 tool calls, 21 s. Passed 5 of 5.
Merge duplicated code
Replace three copies of a send-email function with one, without changing any message sent.
Claude Code: $0.09 (worst $0.09). 29k input tokens (worst 29k), 1,975 output, 2 tool calls, 21 s. Passed 5 of 5.
Codex: $0.25 est. (worst $0.26). 64k input tokens (worst 82k), 1,186 output, 4 tool calls, 32 s. Passed 5 of 5.
Find two bugs
A failing report test whose real causes are a CSV parser that drops quoted lines and a month-key format.
Claude Code: $0.10 (worst $0.11). 40k input tokens (worst 51k), 1,821 output, 3 tool calls, 21 s. Passed 5 of 5.
Codex: $0.15 est. (worst $0.21). 64k input tokens (worst 79k), 767 output, 5 tool calls, 27 s. Passed 5 of 5.
Sales analysis
Five metrics from 5,000 orders in a CSV: revenue by region, top products, refund rate, best month, average order.
Claude Code: $0.11 (worst $0.12). 49k input tokens (worst 50k), 2,542 output, 4 tool calls, 26 s. Passed 5 of 5.
Codex: $0.17 est. (worst $0.18). 61k input tokens (worst 70k), 736 output, 3 tool calls, 25 s. Passed 5 of 5.
Data cleanup
Clean 360 customer rows: trim, fix case, three date formats, phone numbers, drop blanks, resolve 50 duplicates.
Claude Code: $0.13 (worst $0.21). 54k input tokens (worst 121k), 2,920 output, 4 tool calls, 29 s. Passed 5 of 5.
Codex: $0.15 est. (worst $0.17). 62k input tokens (worst 70k), 845 output, 3 tool calls, 26 s. Passed 5 of 5.
Add a feature with tests
Add capped discount codes (like SAVE20CAP50) to a small shop package, with new tests, checked by hidden tests.
Claude Code: $0.14 (worst $0.16). 43k input tokens (worst 56k), 3,549 output, 3 tool calls, 33 s. Passed 5 of 5.
Codex: $0.21 est. (worst $0.28). 64k input tokens (worst 72k), 1,468 output, 4 tool calls, 37 s. Passed 5 of 5.
Email triage
Categorize 200 short support emails (about 20 words each, 3,900 words in all) into five categories and flag the 15 that are urgent.
Claude Code: $0.18 (worst $0.20). 57k input tokens (worst 58k), 2,414 output, 3 tool calls, 28 s. Passed 5 of 5.
Codex: $0.31 est. (worst $0.35). 113k input tokens (worst 124k), 836 output, 4 tool calls, 31 s. Passed 5 of 5.
Refund decisions
Decide 14 short refund requests under a one-page written policy with proration, a fee, an approval threshold and an unknown sender.
Claude Code: $0.19 (worst $0.20). 70k input tokens (worst 85k), 4,591 output, 5 tool calls, 45 s. Passed 5 of 5.
Codex: $0.22 est. (worst $0.25). 80k input tokens (worst 80k), 1,039 output, 4 tool calls, 36 s. Passed 5 of 5.
Questions across 60 notes
Answer five questions across 60 meeting notes (6,200 words) where decisions change over time.
Claude Code: $0.22 (worst $0.23). 102k input tokens (worst 129k), 2,250 output, 5 tool calls, 25 s. Passed 5 of 5.
Codex: $0.27 est. (worst $0.35). 98k input tokens (worst 119k), 559 output, 4 tool calls, 25 s. Passed 5 of 5.
Long contract
Pull 12 current terms from a 16,000-word contract with a later amendment and an outdated exhibit.
Claude Code: $0.35 (worst $0.36). 148k input tokens (worst 169k), 1,781 output, 5 tool calls, 21 s. Passed 5 of 5.
Codex: $0.34 est. (worst $0.44). 151k input tokens (worst 166k), 976 output, 6 tool calls, 36 s. Passed 5 of 5.

How it was measured

Every job uses made-up data and has a fixed pass check. Code jobs also run hidden tests the agent never sees. A reference solution passes every check, and an untouched copy fails them.
Each run starts from a fresh copy of the job. Both agents ran without personal settings, memory, skills or extra connectors, with only their built-in file and shell tools and no web access.
Default reasoning effort, one Mac, one day.

Limits

The jobs are small and synthetic. Real work is larger and messier.
Both agents passed everything, so this shows the cost of work done right, not where agents fail.
One model per agent and five runs per job. Prices and models change, so tokens are the lasting measure.

Common questions

How much does an AI agent cost per task?

In these runs, $0.09 to $0.35 per job at API prices, for jobs that took an agent 20 to 45 seconds. The biggest driver was how much the agent had to read: a 16,000-word contract cost about four times as much as a code change. Your own cost depends on your model, your prices and how much your agent reads.

Why does a tiny task still use tens of thousands of tokens?

Every step, the agent re-reads its own instructions, its tool descriptions and the conversation so far. Here that floor was about 29k input tokens for Claude Code and 60k for Codex, even for the smallest jobs. Most of it is cached, which providers charge far less for.

Is Claude Code or Codex cheaper?

On these jobs Claude Code was cheaper on most small jobs and the two were about even on the longest documents. Codex read more input but wrote about a third as much output. Over all 50 runs each, that was $7.80 for Claude Code and an estimated $11.17 for Codex. Different models, settings or prices would change this, and both got every job right.

Can I use these numbers in a business case?

As a starting estimate, yes, with the limits below. Measure your own agent once it runs, because real work is larger and messier. The one-page agent summary template has a cost section where these numbers fit.

What everyday AI agent jobs cost

How to estimate your own job

Results per job

SQL questions

Merge duplicated code

Find two bugs

Sales analysis

Data cleanup

Add a feature with tests

Email triage

Refund decisions

Questions across 60 notes

Long contract