Claude AI Agent Development, Built Into the Systems You Already Run

We pick the model for each job. Claude does well on long documents and on many tool calls in a row, and every current Claude model reads up to 1 million tokens at once, which suits contracts, catalogs and long email threads. GPT or Gemini can win on a given task. During the demo we run your real examples on two models and show you the results side by side before you commit.

It depends on the step. Anthropic’s models overview, checked on 29 October 2026, lists four current models: Haiku 5.5 for high-volume, latency-sensitive tasks such as classification, extraction and routing, Sonnet 5.5 as the best combination of speed and intelligence, Opus 5.5 for long-running agentic coding and knowledge work, and Fable 5.1 for demanding reasoning and long-horizon agentic work. Many agents use two: a fast model to sort incoming work and a stronger one for the hard cases. We name the exact model ID in your build document.

Usually, yes. Anthropic lists Claude models on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Running Claude in the cloud you already pay for keeps billing, access control and security review in one place, and Bedrock offers regional endpoints that keep data routing inside a region. We confirm which models your cloud offers in your region before we quote.

Yes. We put every model call behind one small layer in your code, so a switch means changing a setting and re-running the test set we built with you. The test set matters more than the switch, because it tells you whether the new model still gets your edge cases right. Prompts usually need some tuning after a move, and we quote that work before we start.

The chat app answers questions. An agent does work inside your systems: it reads a purchase order, checks stock in NetSuite, drafts the order and waits for approval. That takes code that calls the Claude API, tools that read and write your systems, permission rules, logs and tests. Anthropic also asks that products built on its Agent SDK use API keys rather than a claude.ai sign-in.

The Claude Agent SDK is Anthropic’s library for building agents in Python or TypeScript. It gives an agent the same tool loop that powers Claude Code, plus built-in tools, hooks, subagents, sessions, permissions and MCP connections. We use it when the agent works with files, runs code or needs long sessions. For a narrow job, such as reading one type of email, a direct Claude API call with a few tools is simpler to test.

MCP, the Model Context Protocol, is an open standard for connecting AI apps to outside systems. An MCP server is a small program that exposes chosen actions from your ERP or CRM, such as look up order or create draft quote. You need one when more than one AI tool should use the same data, or when staff want Claude on their desktop to reach your systems. It exposes only the actions you approve.

Through tool use, which Anthropic also calls function calling. We describe each tool to Claude with a name, a description and an input format. When Claude decides it needs a tool, it returns a structured request. Our code runs that request against your system, checks the result and sends it back. Claude never touches your database directly. Our code sits in the middle and can refuse a request.

Anything with an API. We connect agents to NetSuite, Odoo, SAP Business One, Salesforce, HubSpot, Zendesk, Gorgias, Shopify, WooCommerce, QuickBooks and Xero, plus email, Slack and punchout catalogs. If a system has no API, we tell you on the first call, because the agent may then need file exports or a person in the loop, and that changes the plan and the price.

Yes. A common job: read an inbound email or form, find or create the contact, log the activity, score the lead against your rules and draft a reply for a rep to approve. The agent works through the CRM’s own API, using an account that has only the permissions it needs. Every write is logged, so your admin can see what changed and why.

Yes. Punchout lets a buyer’s procurement system, such as Coupa or SAP Ariba, open your catalog and send back a cart. A Claude agent can read the purchase order that follows, match each line to your SKUs and price list, and draft the sales order in your ERP. Lines that do not match go to a person with the reason, so nothing is guessed.

A pilot on one narrow workflow usually takes two to four weeks. A production agent with permissions, logs, approvals and monitoring usually takes six to twelve weeks. Most of that time goes into connecting your systems and testing edge cases. The Claude Agent SDK ships hooks, sessions and permission rules, so we do not write those by hand. The demo on your own data comes first, before any contract, so you judge the agent before the clock starts.

Not by default. Anthropic’s privacy center says it does not use inputs or outputs from its commercial products, which include the API, to train its models by default. Training can happen if you opt in, for example by sending feedback. We still send the model only the fields each step needs, and we keep customer details out of prompts wherever the job allows it.

Anthropic says it deletes API inputs and outputs on its backend within 30 days by default. Some cases differ, such as content flagged in a policy review or a signed zero data retention agreement. If your security team needs a different term, that agreement sits between you and Anthropic or your cloud provider, so when you want that agreement in your name, we set the agent up on an account in your company’s name.

Yours, if you want it. We set the agent up on your own Anthropic API key or your own Bedrock, Google Cloud or Foundry account, so usage bills go straight to you at cost, with no markup from us. You can see every request in your own console. If you prefer, we manage the account for you and pass the usage cost through without a markup. Either way, we keep managing the servers, AI models, APIs and maintenance.

Three layers. First, the agent can only call the tools we give it, and each tool checks its inputs. Second, anything that moves money, changes a price or emails a customer waits for a person to approve it. Third, every step is logged, and we test the agent against a set of your real past cases before launch and again after every change.

Any language model can. So our code checks the output against the source. When an agent pulls a name, an amount or a SKU from a document, the code confirms that value exists in the document or in your system before saving it. Our Washington Law Group agent runs this check on every name it extracts, and it has caught and dropped invented names in real runs.

As a market reference, development firm ProductCrafters puts 2026 AI agent builds at about $5,000 to $180,000 or more. The number of systems the agent reads and writes to moves the price far more than the choice of Claude. FactoryJet quotes a fixed price in writing for your scope after one scoping call, and you see the agent working on your own data first.

Rarely by much. The build cost sits in connecting your systems, writing approval rules and testing edge cases, and that work is the same whichever model sits in the middle. Model choice changes your usage bill, which goes to your own account. For each step we pick the smallest model that passes your test set, and we show you the test results behind that choice.

Yes, from day one. The code lives in your own Git repository, along with the prompts, the tool definitions, the test set and any MCP server we write. There is no license fee from us. You can keep us on for support, hand the work to your own team or bring in another firm. Claude usage itself is between you and Anthropic or your cloud provider.

Defects are fixed inside the build price for a cover period we agree before work starts. After that, support is optional: each change is quoted and approved before we start, with no monthly minimum. Agents drift when your prices, policies or data change, and Anthropic retires older models on a published schedule, so we can watch accuracy and move you to a newer model when one is due.

Yes, please do. Email it to connect@factoryjet.com, or attach it at the second step of the form on this page. A useful brief names the job, the systems it touches, a rough weekly volume and who checks the work today. Bhavesh, our founder, reads every brief and replies within 24 hours with questions or a time for a scoping call.

Yes. We work with clients in the US, the UK and other countries. If your data has to stay in one region, we check which Claude endpoints your cloud offers there before we design the agent, because that can decide whether the agent runs on the Claude API or inside Bedrock, Google Cloud or Foundry.