Building Claude Agents: What a Consulting Engagement Looks Like

You know Claude can do more than answer questions. You’ve read about AI agents replacing part-time hires and seen Anthropic’s Claude for Small Business workflows in action. But going from “this could work for us” to a production agent that reliably handles real business tasks is a different challenge. Most companies get stuck at that transition point, not because the technology is lacking, but because they don’t have a clear path from idea to deployment.

This post walks through what a Claude agent consulting engagement actually looks like, phase by phase.

Phase 1: Discovery and Workflow Mapping (Weeks 1-2)

The engagement starts with understanding your operations, not the technology. A consulting team should spend the first two weeks inside your workflows before writing a single line of code.

What happens during discovery:

• Process interviews with the people who actually do the work. Not just managers describing the ideal process, but the frontline staff who know where it breaks down. If you’re automating accounts receivable follow-up, the person currently sending collection emails knows which customers need different approaches and which steps are actually bottlenecks.
• Data audit. What systems hold the data the agent will need? Where does it live (QuickBooks, SharePoint, a shared inbox, a spreadsheet someone emails around)? What format is it in? What access controls exist? This determines whether an agent can be built quickly or whether data infrastructure work needs to happen first.
• Success criteria definition. Before building anything, you define what “working” means in measurable terms. That might be “reduce average collection time from 45 days to 30 days” or “handle 80% of tier-1 support tickets without human escalation.” Vague goals like “improve efficiency” are not useful here.

The output is a workflow map that documents the current process, identifies the steps an agent will handle, and defines the handoff points where humans stay involved.

Phase 2: Agent Architecture Design (Weeks 2-3)

With the workflow mapped, the consulting team designs how the agent will work. This is where Anthropic’s tooling matters.

Claude’s agent capabilities break into three categories:

• Tool use allows Claude to call external APIs, query databases, and take actions in other systems. An agent that processes invoices uses tool use to read data from your accounting system, match it against purchase orders, and flag discrepancies.
• Computer use lets Claude interact with applications the way a human would, clicking through interfaces and filling in forms. This matters when your business runs on legacy software that doesn’t have an API.
• Multi-agent orchestration breaks complex workflows into manageable pieces. Instead of one agent that does everything, you design specialized agents that each handle one part of the process. Anthropic’s Claude agent SDK provides the framework for coordinating these agents.

Common orchestration patterns:

• Single agent with tools for straightforward workflows like data entry or document summarization.
• Router pattern where one agent receives all incoming work and delegates to specialized agents based on the task type.
• Pipeline pattern where agents process work sequentially, each handling a specific step before passing results to the next.
• Supervisor pattern where a coordinating agent monitors subordinate agents, handles exceptions, and decides when human review is needed.

The architecture phase produces a design document specifying which pattern fits your workflow, what tools the agents need, how they communicate, and what guardrails prevent unauthorized actions.

Phase 3: Build and Iterate (Weeks 3-6)

Development uses Claude Code, Anthropic’s development environment for building agent systems. The build phase is iterative, not waterfall.

A typical build cycle:

1. Core agent logic first. Build the primary agent that handles the most common path through your workflow. For an A/R automation agent, that means invoice reading, matching, and routing for standard cases.
2. Tool integrations. Connect the agent to your actual systems. Connecting Claude to QuickBooks, your CRM, or your ERP requires building tool definitions that translate between the agent’s understanding and your system’s API. This is where most of the complexity lives.
3. Edge case handling. Every workflow has exceptions: invoices with missing PO numbers, emails in unexpected formats, customers with special billing arrangements. The agent needs explicit logic for handling these, either resolving them or escalating to a human.
4. Guardrails and safety. A financial agent should never approve a payment above a certain threshold without human sign-off. A customer-facing agent should never share internal pricing. These constraints are built into the agent’s system prompt and tool permissions.

The consulting team should demo progress weekly. You review the agent’s behavior on real data from your business, not generic test cases. If the agent handles 90% of your invoices correctly but misreads invoices from one particular vendor, that feedback needs to surface early.

Phase 4: Testing and Deployment (Weeks 5-8)

Testing an AI agent is different from testing traditional software. A conventional application either produces the correct output or it doesn’t. An agent’s output involves judgment.

Testing layers:

• Unit tests verify that individual tool integrations work. Can the agent read data from your accounting system? Does the email function work? These are pass/fail.
• Scenario tests run the agent through complete workflows using historical data. Take the last 200 invoices your team processed manually and run them through the agent. Compare the agent’s decisions to what your team actually did.
• Adversarial tests try to break the agent. What happens with garbage data? What if someone tries to bypass guardrails?
• Human evaluation brings your staff into the process. The people who will work alongside the agent review its outputs and flag anything that doesn’t match their judgment.

Deployment follows a gradual rollout:

• Shadow mode. The agent processes real work but doesn’t take action. Your team reviews its recommendations side-by-side with their manual process.
• Supervised mode. The agent handles 20-30% of real work with human approval required for every action.
• Autonomous mode with exceptions. The agent handles routine cases independently and escalates exceptions. This is the target operating mode for most workflows.

Each stage should last one to two weeks. Rushing from shadow mode to full autonomy is how companies end up with agents sending incorrect invoices to real customers.

Timeline and Milestone Framework

A single-workflow agent engagement typically runs 8 to 10 weeks from kickoff to autonomous operation. Multi-agent systems or workflows requiring complex legacy integrations can extend to 12 to 16 weeks.

Ask any consulting partner to break their proposal into these phases with clear deliverables at each stage. If they can’t, they haven’t done this before.

Choosing the Right Partner

Not every AI services provider has experience building Claude agent systems. When evaluating partners, ask:

• Have you built production agents using Anthropic’s tools? General AI experience is not the same as Claude-specific agent development. The tool use patterns, prompt engineering, and orchestration approaches are specific to Anthropic’s platform.
• What’s your approach to testing AI outputs? If the answer is only unit tests, they’re missing the evaluation work that determines whether an agent is ready for production.
• How do you handle the handoff? A good engagement doesn’t create permanent dependency. Your team should understand how the agent works, how to monitor it, and how to make basic adjustments without calling the consulting team back.

The first agent project is as much about building internal capability as automating a specific workflow. Your team learns how to work with AI agents, how to provide feedback that improves them, and how to identify other workflows that are candidates for automation. The AI governance framework you put in place for the first agent applies to every agent you build after it. That organizational learning is worth as much as the agent itself.