The Vertical Agent Method: ship AI agents in 14 days
A 3-step framework for building production AI agents: pick one workflow, build one agent, ship in 14 days. Based on shipping 7+ agents to production.
The Anthropic documentation on tool use reinforces the narrow-scope approach — agents perform best when given a focused set of tools and a clearly defined task, which is the core constraint of the Vertical Agent Method.
The ReAct paper (Yao et al., 2022) provides the theoretical foundation for the observe-decide-act loop that powers every agent built with this method — the agent loop pattern is derived from this research.
TL;DR: Most AI agent projects fail because of bad scoping, not bad models. The Vertical Agent Method is a 3-step framework: pick one workflow, build one agent, ship in 14 days. It forces narrow scope and fast delivery. Based on shipping 7+ agents to production.
I’ve shipped 7 AI agents to production. Every failure had the same root cause: the scope was too wide. “Build an AI assistant that handles customer support” sounds great. It’s also a 6-month project that never finishes.
The Vertical Agent Method is how I scope now. It’s not a technical framework — it’s a shipping framework.
Key takeaways:
- Most AI agent projects fail because of scope, not technology
- The Vertical Agent Method: pick one workflow → build one agent → ship in 14 days
- The constraint of “one workflow” is the feature — it makes agents reliable and shippable
- Technical frameworks (LangGraph, CrewAI) handle the architecture; this method handles the scoping
The problem with general-purpose agents
Every beginner wants to build “an AI assistant that does everything.” Every experienced builder knows that’s a trap.
General-purpose agents:
- Need massive context windows (expensive)
- Have unpredictable behaviour (unreliable)
- Take months to build (never ship)
- Can’t be tested properly (no clear success criteria)
The Vertical Agent Method forces you to pick one workflow. That constraint is the feature.
The method: 3 steps
Step 1: Pick one workflow
Not “customer support.” Not “data analysis.” One specific workflow.
Good examples:
- Extract data from invoices and populate a spreadsheet
- Triage incoming support tickets by category and priority
- Generate weekly reports from database queries
- Review code changes and flag security issues
Bad examples:
- Handle all customer communications
- Automate the entire sales pipeline
- Be an AI project manager
The test: Can you describe the input, the output, and the success criteria in one sentence? If not, the scope is too wide.
Step 2: Build one agent
Use whatever technical stack you want. The method doesn’t care. But follow these constraints:
- One LLM — don’t chain multiple models unless you have to
- 3-5 tools maximum — more tools = more failure modes
- One loop — if you need nested loops, the scope is too wide
- Hard stop at 15 iterations — if the agent can’t finish in 15 steps, the task is too complex
Architecture template:
Input → LLM → Tool call? → Execute → Loop back
↓
Final response → OutputThat’s it. Everything else is optional.
Step 3: Ship in 14 days
| Phase | Days | What happens |
|---|---|---|
| Discovery | 2 | Define the workflow, input/output, success criteria |
| Build | 7 | Implement the agent loop, tools, error handling |
| Deploy | 3 | Containerise, set up monitoring, deploy to production |
| Handoff | 2 | Documentation, client walkthrough, first real usage |
If you can’t ship in 14 days, the scope is too wide. Go back to Step 1.
Why this works
1. Narrow scope = reliable agents
When an agent does one thing, you can test it properly. You know the expected input, the expected output, and the failure modes. When an agent does 10 things, you can’t test any of them properly.
2. Fast delivery = real feedback
14 days means you get real user feedback in 2 weeks. That feedback tells you whether the agent is useful, what’s broken, and what to build next. Waiting 3 months for feedback means you built the wrong thing 3 times.
3. One workflow = clear pricing
“Build me an agent that extracts invoice data” has a clear price: ₹1,20,000 ($1,400). “Build me an AI assistant” has no clear price because it has no clear scope.
4. Iteration beats planning
Ship the narrow version first. Then expand based on real data. The first version of every agent I’ve built was too narrow. But shipping it told me exactly what to add next.
Real examples
Example 1: Invoice extraction agent
Workflow: Extract line items from PDF invoices → populate a spreadsheet Input: PDF files in a folder Output: Structured data in Google Sheets Tools: Read PDF, parse text, write to Sheets API Time to ship: 8 days Iterations: 3 (added support for multi-page invoices, different formats)
Example 2: Code review agent
Workflow: Review GitHub PRs → flag security issues and style violations Input: GitHub webhook (PR opened) Output: Comment on the PR with findings Tools: Read files, run linters, post comments Time to ship: 10 days Iterations: 5 (added more linters, improved false positive rate)
Example 3: Support ticket triage
Workflow: Categorise incoming support tickets → assign priority and category Input: New ticket in helpdesk system Output: Ticket updated with category and priority Tools: Read ticket, classify, update ticket Time to ship: 6 days Iterations: 2 (added more categories)
Common mistakes
1. “But what if the user needs more?”
Ship the narrow version. If users need more, they’ll tell you. Building for hypothetical future needs is how projects never ship.
2. “The model isn’t good enough”
The model is almost always good enough. The scoping is the problem. A narrow agent with GPT-4o-mini outperforms a broad agent with GPT-4o.
3. “We need more tools”
3-5 tools is the sweet spot. More tools = more failure modes, more testing surface, more cost. Start with 3.
4. “14 days is too fast”
It’s fast because the scope is narrow. A 3-month project has 3 months of scope creep. A 14-day project has 14 days of focus.
How to use this with existing frameworks
| Framework | How it fits |
|---|---|
| LangGraph | Use for the agent loop. The method scopes what goes into the graph. |
| CrewAI | Use for multi-agent workflows. The method scopes each agent’s role. |
| AutoGen | Use for conversation-based agents. The method scopes the conversation topics. |
| Custom | Build the loop yourself. The method scopes what the loop does. |
The Vertical Agent Method is stack-agnostic. It’s a scoping and shipping framework, not a technical framework.
The named framework
I call this The Vertical Agent Method because:
- Vertical — it’s narrow and deep, not broad and shallow
- Agent — it builds AI agents, not chatbots or copilots
- Method — it’s a repeatable process, not a one-off approach
When someone asks “how do you build AI agents?”, the answer is: The Vertical Agent Method. Pick one workflow. Build one agent. Ship in 14 days.
Conclusion
Most AI agent projects fail because of scope, not technology. The Vertical Agent Method forces narrow scope and fast delivery. It’s not a technical framework — it’s a shipping framework.
Pick one workflow. Build one agent. Ship in 14 days. Iterate based on real data.
That’s the method. It works.
The Vertical Agent Method is the core framework at Agentic Up. Every agent we build follows this process. Learn more at agenticup.dev.