How to Create an AI Agent: A Tested Builder's Guide (2026)

You don't create an AI agent in 2026 by picking a framework and writing code from scratch. You pick a path based on your situation, then use the platform that fits that path.

There are four real paths. Each ships in hours-to-weeks, has a predictable cost shape, and uses platforms we've tested in production. This guide walks through all four, with the honest verdict on when each is right.

If you only read one paragraph: non-developers should start with Lindy or Manus AI. Developers with a Claude subscription should start with Claude Code. Technical teams with data control or scale concerns should pick OpenClaw, Hermes, or Paperclip. Everything else in this article is helping you pick between those.

Step zero: define what the agent actually does

The single biggest reason agent projects fail isn't model choice, platform choice, or code quality. It's that the builder didn't write down a clear definition of success before starting. The agent ships, runs, produces output, and nobody can agree whether it worked.

Before opening any platform, answer three questions in writing:

1. What does the agent do in one sentence? Not "improve customer support" — that's a goal. "Read incoming support tickets, classify them by category and urgency, draft a first response, route to the right queue" — that's an agent.
2. What inputs does it have access to? Email inbox, CRM, knowledge base, internal docs, calendar, web. Be specific. Every input is a permission decision and an integration cost.
3. How will you know it worked? "The first-response time drops below 15 minutes" is testable. "It feels faster" isn't. Write the metric.

If the answers feel obvious, write them anyway. The discipline of putting them in a doc surfaces the assumptions that would otherwise blow up the project in week three.

The picker is a five-question version of this exercise if you want to walk through it interactively.

Path 1 — No-code (the right pick for most non-developers)

Best for: sales operators, support managers, marketing teams, founders without a developer on call. Time to first agent: 1–4 hours. Monthly cost: $39–$199 for the platform, plus $5–$50 in model costs for typical volumes.

The platforms worth considering:

• Lindy — the most operator-friendly platform in 2026. Pre-built templates for sales follow-up, support triage, meeting prep, and lead qualification ship working in under an hour. Plans from $49/month. The right default for most non-technical builders.
• Manus AI — autonomous browser-based agents. Best for research and exploration tasks where the agent needs to browse the web and produce structured deliverables. ~$39/month, free tier available.
• Stack AI — the strongest pick when your use case is document Q&A. Ingest internal SOPs, contracts, product specs, Notion wikis; the agent answers questions against them in plain English. From $199/month for serious team use.

How to actually do it on Lindy (the most common starting point):

1. Sign up for a free Lindy account
2. Pick the template closest to your job (e.g., "Sales follow-up assistant")
3. Connect the tools the agent needs — Gmail, Slack, HubSpot
4. Customize the prompt with your company's tone and rules
5. Test on five real examples; check the output
6. Deploy

The whole loop is hours, not days. The trade-off: you're on their infrastructure, their pricing, and their integration list. For standard B2B workflows with non-sensitive data, that trade-off is worth it.

When this path is the wrong pick: you need customisation beyond what templates allow, you handle regulated data, or you'll exceed Lindy's volume tier and want cheaper unit economics. In that case, jump to Path 2 or Path 3.

Path 2 — Low-code (the right pick for technical operators)

Best for: ops teams with at least one developer, builders who outgrew no-code limitations, teams who want self-host as an option. Time to first agent: 1–3 days. Monthly cost: $0 (self-hosted) to $199/month (cloud), plus $20–$200 in model costs.

The platforms worth considering:

• n8n — the best workflow automation platform for teams with a developer. 400+ integrations, self-hostable, native AI agent layer. Free open-source self-hosted, $24/month for cloud. The right default when no-code SaaS hits a ceiling.
• Relevance AI — sits between Lindy (pure no-code) and n8n (developer-required). Strongest for outbound research and sales intelligence workflows where the logic is complex but the stakes don't justify a custom engineering build. From $19/month.

How to build a customer-support deflection agent on n8n:

1. Spin up n8n — either Docker on a $5/month VPS, or sign up for n8n cloud
2. Wire the trigger (incoming Intercom message, Zendesk ticket, etc.)
3. Add a knowledge base node that queries your docs
4. Add an AI agent node — pick Claude Sonnet 4.6 as the model (the cost calculator shows why)
5. Add conditional logic: if the agent's confidence is high, send the reply; if low, escalate to a human queue
6. Add observability (log every action) and a per-day budget cap

The first version is usually working in a day. Tuning the prompt and edge cases takes the rest of the week.

When this path is the wrong pick: you're a non-technical operator with no developer access, or you need pure agentic loops without external system integrations (use Path 3 or a dedicated harness instead).

Path 3 — Developer / custom (the right pick if you can code)

Best for: developers, technical founders, engineering teams shipping production agents. Time to first agent: 3–10 days. Monthly cost: $0 platform, $20–$500 in model costs depending on volume.

The real options in 2026:

• Claude Code — the fastest path from idea to working agent if you have a Claude subscription. Included in Claude Pro ($20/month). The CLI runs on your machine, reads your codebase, writes the agent, runs it. Best default for developers.
• Cursor — VS Code fork with multi-model support (Claude, GPT, Gemini). Stronger if you want to switch between models mid-session or you prefer an IDE-first workflow. From $20/month.
• OpenAI Codex — OpenAI's terminal coding agent, 3M weekly active users. The right pick if you're committed to GPT-5+ models. $20/month base plus usage credits.
• Amp — Sourcegraph's agentic coding tool, strongest if your team already runs Sourcegraph Code Search. Free tier available, paid from $19/month.
• Kilo Code — open-source coding agent for VS Code, JetBrains, and CLI. The only mature option for JetBrains users wanting agentic workflows. BYOK pricing.
• Roo Code — Cline-lineage fork with opinionated mode-based workflows (Code, Architect, Ask). The right pick if you want stronger guard-rails on what the agent does.
• OpenHands — autonomous coding agent that completes full engineering tasks (PR reviews, CVE fixes, legacy migrations). Different category from inline assistants — task-complete architecture. 65k GitHub stars, open source.
• OpenAI Agents SDK — code-level framework for building custom agents in TypeScript or Python. Less polished than Anthropic's tooling for agentic workflows; pay only for the API tokens consumed.
• Anthropic Agent SDK — the closest to "build a custom agent from first principles" while still getting Anthropic's tool-use, computer-use, and routing infrastructure. Pay only for the API tokens consumed.

The coding agents shortlist has the full side-by-side if you're weighing between several of these.

How to build a research agent with Claude Code:

1. Install Claude Code (one CLI command)
2. Create a CLAUDE.md describing your agent's job, constraints, and success criteria
3. Tell Claude Code to scaffold the agent — it writes the code, the API integrations, the prompt, the error handling
4. Test on real examples, iterate on the prompt
5. Add observability via OpenTelemetry or LangSmith
6. Deploy to your hosting layer (Vercel, Modal, your own infra)

The advantage over no-code/low-code: you can do anything. The trade-off: you're responsible for the operational infrastructure that platforms ship by default — observability, budget caps, audit logs, error recovery.

When this path is the wrong pick: you're not comfortable in a terminal, your team doesn't have developer capacity, or the workflow is simple enough that Lindy or n8n handles it without you reinventing wheels.

Path 4 — Open-source / self-hosted (the right pick for data control)

Best for: technical teams with strict data residency, compliance requirements, or scale where unit economics of SaaS platforms break down. Time to first agent: 2–7 days. Monthly cost: $5–$50 for hosting, plus model API costs.

The three real options:

• OpenClaw — 365k GitHub stars, the most mature open-source agent harness. Install it locally or on a server, connect to messaging platforms (WhatsApp, Slack, Telegram), use any AI model. MIT licence.
• Hermes — Nous Research's server-deployed harness. Builds institutional memory from experience. Strong for sustained workflows where the agent gets better over time. MIT.
• Paperclip — the orchestration layer for teams running multiple agents. Org charts, budget limits, approval gates, audit logs. MIT. Pair with OpenClaw or any agent runtime underneath.

How to set up OpenClaw on a server:

1. Spin up a VPS or use an existing server with Docker
2. Pull the OpenClaw image, configure the messaging connector you want
3. Add API keys for the model(s) you'll use
4. Configure the agent's persona, tools, and budget caps
5. Connect to your data sources
6. Run

The community is the biggest advantage — 365k stars means problems you'll hit have been solved publicly, often with documentation.

When this path is the wrong pick: you don't have technical capacity to operate a server, your data isn't sensitive enough to need self-hosting, or you'd be reinventing operational infrastructure that Lindy or Paperclip cloud handles.

What every path needs (regardless of which you pick)

Four things every production agent needs, on every path:

Model selection. Default to Claude Sonnet 4.6 for ~90% of workloads — best price-performance balance in 2026 ($3 input / $15 output per 1M tokens). Use Claude Opus 4.8 or GPT-5.5 only when the task genuinely needs frontier reasoning. Use Claude Haiku 4.5 or Gemini 2.5 Flash for high-volume cheap work like classification and routing. The cost calculator shows the side-by-side math for any specific workload.

Observability. Log every agent action — the prompt, the model response, tokens used, latency, success or failure. Ship those logs somewhere queryable. Without observability, your first cost leak takes weeks to catch. Our observability guide covers what to instrument.

Guardrails. Budget caps, action boundaries, approval gates for irreversible actions. AI agents don't fix bad architecture, they amplify it. Our guardrails article covers the seven every production deployment needs.

A test set before you ship. Five to ten real-world examples with expected outputs, run before every deploy. Even informal testing prevents the "looked good in demo, broken in production" problem.

The cost reality

What it actually costs to run an AI agent depends on three things: the platform fee (if any), the model API cost per task, and the volume.

Honest numbers for a sales follow-up agent running 1,000 tasks per month:

Path	Platform	Model	Monthly total
No-code	Lindy ($49)	Claude Sonnet via Lindy	$80–$150
Low-code	n8n cloud ($24)	Claude Sonnet (BYOK)	$40–$60
Low-code	n8n self-hosted ($5 VPS)	Claude Sonnet (BYOK)	$20–$40
Developer	Claude Code ($20 Pro)	Claude Sonnet	$50–$100
Open-source	OpenClaw ($5 VPS)	Claude Sonnet (BYOK)	$20–$40

The article "How much does it cost to build an AI agent" has the full breakdown across more workflows.

The five mistakes builders make most often

1. Skipping step zero. Writing the success criteria before opening any platform is the highest-leverage hour in the whole project. Most teams skip it because it feels like overhead. It isn't.
2. Defaulting to the frontier model. Most production agents run fine on Claude Sonnet 4.6 or Haiku 4.5. Defaulting to Opus 4.8 or GPT-5.5 on every call burns money for marginal quality gains on most workloads.
3. No observability. Shipping without logs means you learn about failures from users, not metrics. The instrumentation is cheap; the cost of operating blind is high.
4. No budget cap. A runaway agent loop has cost real teams real money. Cap per-day spend at the platform level before the first deploy.
5. Picking the wrong path. Non-developers fighting through n8n setup, or developers using Lindy for something that needs full code control — both fail because the chosen path doesn't fit the team's capacity. Be honest about your starting point.

What to read next

The 2026 AI agent shortlist is the platform-by-platform breakdown across every category mentioned here. The cost calculator lets you size the model bill for your specific workload across 25 frontier and value-tier models. The agent picker is a five-question version of step zero — answer five questions, get a platform recommendation matched to your situation.

If you're stuck deciding between paths, the picker is the fastest way to break the tie.