An AI agent is software where a language model decides its own next steps and uses tools, such as a CRM, a spreadsheet or an email draft, in a loop until a task is done. Agents can take over lead intake, quote drafts, support triage, invoice matching and report drafting. A person should still approve anything that spends money, contacts a customer or changes a record you cannot easily restore.
Devvista is a software development and growth agency that builds automation and AI agents for business owners. This post defines the term using Anthropic's own engineering guidance, sorts the jobs agents suit from those they do not, and explains the guardrails and how we scope a first agent.
01 What an AI Agent Actually Is
Anthropic's engineering team separates two things that are often blurred. In "Building effective agents" it defines workflows as systems where language models and tools are orchestrated through predefined code paths, and agents as systems where the model dynamically directs its own process and tool use. Agents are typically just models using tools based on environmental feedback in a loop.
The Claude Agent SDK documentation describes an agent as an application that completes a task by planning its own steps and calling tools. The SDK gives developers the same tools, agent loop and context management that power Claude Code, in Python and TypeScript.
Our plain-English version (ours, not Anthropic's): an agent is a digital assistant that is given a goal and some tools, works out the steps itself, and stops to ask when it needs a person.
The practical difference: a workflow follows the path you drew, an agent chooses the path. That flexibility is the point, and it is also the risk.
02 Workflow or Agent: Which One Does Your Job Need?
Anthropic recommends finding the simplest solution possible and increasing complexity only when needed. For most business jobs, that means a workflow first.
| Workflow automation | AI agent | |
|---|---|---|
| Who decides the steps | You, in advance | The model, as it goes |
| Best for | Well-defined, repeatable tasks | Open-ended tasks where the steps vary case by case |
| Predictability | High: same input, same path | Lower: the path can differ every run |
| Cost and speed | Usually cheaper and faster | Often trades cost and latency for better task performance, per Anthropic |
| Typical failure | A step breaks or an input changes | A wrong decision that compounds across steps |
| Example | New form submission creates a CRM contact and sends a confirmation | Reading a messy request, looking up three systems and drafting a tailored reply |
The same article names five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. Many useful business systems are a workflow with one or two AI steps inside it, with no self-directed agent at all. See our workflow automation service for that side.
03 Business Jobs Agents Can Take On, and Where a Person Stays
This table is our working view of where agents earn their place. Each job is one where reading unstructured text and choosing among tools beats a fixed rule, and each has a point where a human decides.
| Job | What the agent does | Where a person approves |
|---|---|---|
| Lead intake and qualification | Reads a form or email, enriches it, scores fit against your criteria, routes it | Rejecting a lead, or any first message to a prospect |
| Quote drafting | Pulls prices and past quotes, drafts the quote and a cover note | Every quote before it is sent, and any discount |
| Support triage | Classifies a ticket, finds the relevant account and articles, drafts a reply | Refunds, account changes and sensitive cases |
| Invoice matching | Compares an invoice with the purchase order and delivery note and flags mismatches | Payment release and every mismatch |
| Report drafting | Pulls the numbers, writes the first draft with the figures sourced | Final wording and any claim made externally |
| Data entry between systems | Extracts fields from documents and writes them to the right records | Low-confidence extractions and unusual formats |
Anthropic names customer support and coding as good fits, since support agents can pull customer data and take actions through tools, and code is verifiable through automated tests. For a broader list of proven jobs, see our AI automation use cases for companies.
Which of these jobs is eating your team's week? Book a call and tell us the process; projects are quoted after a short call, and you state your budget in the form.
04 What We Have Built: Real Projects, Described Carefully
These four projects are published on our work page, and everything below comes from those pages. Two caveats up front: the clients are confidential where the pages say so, and these systems were built on workflow platforms and AI tools such as Make.com, GoHighLevel, Relevance AI and Claude, not as standalone agents built in code.
- AI lead intake and blueprint automation (Make.com and Claude). A confidential consulting client. A 112-field Typeform feeds GoHighLevel and Airtable without dropping fields, and a Claude prompt of more than 1,500 words generates a business blueprint saved as a Google Doc. Nothing is delivered until internal approval, with a Slack review step and status-based control.
- Lead qualification and follow-up in GoHighLevel. A confidential sales team. Three workflows run on custom fields and conditional logic so a lead cannot move forward without proper qualification, with a monthly follow-up cadence for unresponsive leads.
- Meta ads to CRM automation. A confidential mortgage and finance client in the US and UK. Make.com, Typeform and GoHighLevel connect Meta lead ads to the CRM, avoid duplicate records, and send a qualified-lead event back to Meta through the Conversions API.
- AI credit repair automation for Harold. A multi-agent analysis system built with Relevance AI (17 agents), with Make.com, Airtable and GoHighLevel around it. Dispute letters are generated with human checkpoints, and high-risk cases are flagged for manual review.
The pattern across all four is the same as the rest of this post: AI handles reading and drafting, rules handle sequencing, and a person approves what leaves the building.
05 Guardrails: The Five Controls a Business Agent Needs
Anthropic advises testing extensively in sandboxed environments, adding stopping conditions such as a maximum number of iterations, and letting agents pause for human feedback at checkpoints or when they hit a blocker. In practice we set five controls before anything goes live.
- Human approval points. The agent proposes and a person approves anything that sends a message, moves money or deletes data.
- Logs. Every run records what the agent read, which tools it called and what it decided, so a bad outcome can be traced.
- Permissions. The agent gets the narrowest access that works: read-only first, one system at a time, no access to secrets it does not need. Hooks can block specific actions outright.
- Error handling. What happens when a tool is down, an input is malformed or the model is unsure: retry, skip, or hand to a person.
- Cost and loop limits. A cap on steps and spend per task, because agents trade cost and speed for flexibility.
Connected tools add a risk of their own. Anthropic warns that servers fetching outside content can expose you to prompt injection, so an agent that reads inbound email or web pages should not also hold broad write access. For the rules file behind this, read what a CLAUDE.md file is.
06 How Agents Fail, and What Each Failure Looks Like
| Failure mode | What you see | The fix |
|---|---|---|
| Compounding errors | A small early mistake carries through every later step | Check the output at each step, not only at the end |
| Confident wrong answer | A fluent draft with a wrong price or policy | Source every figure from a system, and require approval |
| Prompt injection | Text inside an email or page tries to redirect the agent | Limit permissions, treat inbound content as data, review actions |
| Runaway loop or cost | The agent keeps trying and keeps spending | Iteration and spend limits, with an alert |
| Silent drift | A source format changes and results quietly worsen | Sample checks every week, plus tests on known cases |
| Nobody owns it | It breaks and no one notices | A named owner, logs and failure alerts |
The security documentation says that no system is completely immune to all attacks and that you are responsible for reviewing what it proposes. We build with that assumption.
07 How We Scope a First Agent in Weeks
We keep the first agent small on purpose. A typical path, which varies by system and is an estimate rather than a promise:
- Pick the job. One process, one owner, one clear definition of a good result.
- Map it. Who does it today, which systems they touch, where it goes wrong, and which steps are fixed rules.
- Decide workflow or agent. Use a plain workflow wherever the steps are known, and an agent only where the path varies.
- Build a read-only version against a copy of the data that drafts but sends nothing.
- Test on past cases and compare with what your team actually did.
- Go live with approval points, logs and limits, then widen the agent's freedom only as the numbers hold up.
In our experience a first agent is scoped in weeks, not months, as an estimate and not a promise, because the scope is one job. We quote after a short call. See our AI agents service and AI automation service for what is included, and our API development work when your systems need connecting first.
08 When Not to Use an Agent
Agents are not always the answer, and Anthropic says to add this kind of complexity only when simpler options fall short.
- The steps never change. A rule-based workflow is cheaper, faster and more predictable.
- A wrong answer is costly and cannot be checked. Legal advice, medical decisions and large payments need a person in charge.
- Your data is a mess. An agent on top of duplicate records and missing fields multiplies the problem. Clean the data first.
- Nobody can say what good looks like. Without a test, you cannot tell if it works.
- The volume is tiny. If a task takes ten minutes a week, build nothing.
If you are not sure, start with our guide to automating business processes with AI, which covers how to find the repetitive work first.
Not sure whether your job needs an agent, a workflow or nothing at all? Book a call or use the contact page. We read what you send, tell you honestly whether the job suits this kind of tool, and quote the work after a short call.
? Frequently Asked Questions
→ Related Resources
- Claude Code for business: what it can automate
- AI automation for business: a plain-English guide
- How to automate business processes with AI
- AI automation use cases for companies
- AI agents service
- AI automation service
- Workflow automation service
- API development service
- Our work and case studies
Local rules change how an agent is built, so each market has its own page: UK, UAE, Australia, Canada, Pakistan and Saudi Arabia.
→ Sources
Opened and checked on 2026-11-11. Product documentation changes often, so check the page itself before you rely on a detail.
- https://www.anthropic.com/engineering/building-effective-agents: Workflows versus agents, agents as LLMs using tools in a loop, start with the simplest solution, cost and latency trade-offs, compounding errors, stopping conditions, human checkpoints, five workflow patterns.
- https://code.claude.com/docs/en/agent-sdk/overview: An agent plans its own steps and calls tools; the Agent SDK provides built-in tools, hooks, subagents, MCP, permissions and sessions in Python and TypeScript.
- https://code.claude.com/docs/en/security: Permission modes, review responsibility, prompt injection safeguards and the point that no protection is complete.
- https://code.claude.com/docs/en/mcp: MCP servers connect agents to tools and data; verify you trust each server because of prompt injection risk.
- https://modelcontextprotocol.io/introduction: MCP is an open-source standard for connecting AI applications to external systems.
- https://code.claude.com/docs/en/hooks-guide: Hooks give deterministic control around agent actions, including blocking a tool call.
- Devvista case studies (/work/): Make.com and Claude lead intake, GoHighLevel lead qualification, Meta ads to CRM automation and the credit-repair multi-agent platform, as described on their published pages.