Custom Claude Agents consulting

A custom agent is code that gives Claude a goal and a defined set of tools, runs the tool calls Claude asks for, and loops until the job is done. A demo takes an afternoon. Production takes the parts demos skip: tool definitions tight enough that Claude does not guess missing values, a permission model for every action that changes data, an evaluation set that runs on each prompt change, logs your security team can read, and a cost model that survives real volume. We scope those parts first, because they drive budget and timeline far more than the first prompt does, and because they are what your team will maintain after launch.

1. 01

   Three ways to build on Claude

   The Messages API gives direct model access: you write the loop, or let the SDK tool runner drive it, and you host everything. The Claude Agent SDK, in Python and TypeScript, packages the Claude Code harness as a library, with built-in file, shell and web tools, hooks, subagents, MCP, permissions and resumable sessions, running in a process you operate. Claude Managed Agents, in beta, is a hosted harness: you define the agent, environment and session through the API, and Anthropic runs the loop in a managed cloud sandbox or in a self-hosted sandbox on your infrastructure. Each step up removes code you would write and adds terms you need to accept.

2. 02

   Where it runs changes what you can use

   The same models are reachable through the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry, and the routes are not equivalent. On Bedrock, AWS operates inference and is the data processor, and Message Batches, the Files API, Agent Skills, the MCP connector and Managed Agents are not supported. Google Cloud (Google managed) and Foundry also lack Message Batches and Managed Agents. Claude Platform on AWS is operated by Anthropic, billed through AWS Marketplace, with feature access typically the same day as the Claude API. List the features your design needs, then pick the route, not the other way around.

3. 03

   Why Abstrakt Solutions

   Most agents our clients ask for touch a CRM, an ERP or both. As a Salesforce partner since 2017 with 108 Salesforce certifications, and a Certified Services Partner in the Claude Partner Network with 12 Claude certifications, we can tell you when a job belongs in Agentforce, when it belongs in a custom Claude agent, and when it needs both connected through MCP or APIs. We build alongside your developers, hand over the eval suite, prompts and runbooks, and quote builds as a fixed fee or in phases after discovery, with consulting at $180 to $225 per hour. Our published Claude projects for Sync Payments and Abstrakt Marketing Group are linked below.

The engagements we run most often on Custom Claude Agents, from first implementation through optimization.

1. 01

   Use case and success criteria

   Owners: business sponsor and technical lead, one to two weeks. We write the agent's job in one paragraph, list the systems it reads and the actions it may take, and define success in measurable terms: accuracy on a labeled sample, latency, cost per task and the error rate the business can live with. Then we settle build or buy, including whether Agentforce or a packaged product already covers it.

2. 02

   Architecture and route

   Owners: architect and security. We pick the build path (Messages API, Agent SDK or Managed Agents) and the hosting route by checking each required feature against the platform tables: batch, Files API, MCP connector, Managed Agents, ZDR and HIPAA eligibility. We define tools with strict schemas and narrowly scoped credentials, and mark which actions need human confirmation. Security signs off on the data flow before code is written.

3. 03

   Eval-first build

   Owner: engineering lead. We assemble a test set that mirrors real inputs, including messy documents, missing fields, out-of-scope requests and injection attempts, and grade with code where possible and a separate model where judgment is needed. The agent is built against that set. Every prompt, tool or model change reruns it in CI, and a regression blocks the merge the way a failing unit test would.

4. 04

   Guardrails and pilot

   Owners: product owner and reviewers, two to four weeks. We add input screening, screening of third-party tool output, structured outputs for anything parsed downstream, and spend and rate limits on a dedicated workspace or cloud budget. A read-only or draft-only pilot runs with a small user group while reviewers grade outputs against the success criteria and log every miss with the input that caused it.

5. 05

   Production and operations

   Owner: named agent owner, ongoing. Write actions are enabled one at a time behind confirmation. Usage, cost and traces flow to your monitoring, and alerts fire on spend thresholds, error spikes and refusal rates. We review quality and cost monthly, rerun evals before any model migration, and keep a written record of who owns prompts, tools, credentials and the eval set.

1. 01

   Choosing the cloud before listing features

   Teams commit to Bedrock, Google Cloud or Foundry and then find Message Batches, the Files API or Managed Agents are not available there. Check the platform feature lists first.

2. 02

   Putting third-party content in the prompt

   Emails, web pages and documents belong in tool_result blocks with their source labeled, never in the system prompt. Screen that content with a small classifier call before the agent can act on it.

3. 03

   Treating auto approval as a human checkpoint

   The Managed Agents docs state that auto is not a human checkpoint: a call judged safe runs before anyone sees it. Set always_ask on any tool a person must approve.

4. 04

   Building the eval set after something breaks

   Without a fixed test set, a prompt tweak or model upgrade can quietly lower accuracy for weeks. Build the set before the agent and rerun it on every change.

Other Claude platforms we implement

• What should an AI acceptable-use policy say, and how do you enforce it in Claude?
• Should you build Claude agents on the Messages API, the Agent SDK, or Managed Agents, or buy one?
• How do you roll out Claude Code to an engineering team safely?
• What does Claude Enterprise cost? Seats, usage, API rates and rollout services
• Claude Enterprise vs. ChatGPT Enterprise vs. Microsoft 365 Copilot: which fits your company?
• How can finance and accounting teams use Claude without weakening controls?
• Claude Team vs. Enterprise: which plan does your company need?
• How do you connect Claude to your company's data? SharePoint, Google Drive, Slack, databases and ERPs
• Does Claude train on your data? Training, retention, zero data retention and where data lives
• How do you roll out Claude company-wide? A phased plan with owners, settings and checkpoints
• Is Claude HIPAA compliant? What the HIPAA-ready offering and BAA actually cover
• How do you measure the ROI of Claude in your company?
• What are Claude Skills, and how do companies build their own?
• What does a Claude implementation partner do, and do you need one?
• Using Claude with Salesforce: where it fits alongside Agentforce