Almost every product roadmap in 2026 has an "AI agents" section. The pattern is usually the same: a list of 8 to 20 agent ideas, generated by founders, sales, customers, and internal teams. The PM tries to prioritize them and gets stuck. The reason for the deadlock is rarely lack of judgment. It is that most teams treat agent ideas as one category when they are actually three architecturally different things with very different costs, timelines, and risk profiles.
Hamza Farooq and Jaya Rajwani articulated this clearly in their Lenny's Newsletter piece. This article distills their three-category framework and adds the prioritization logic for product teams who need to decide what to build first.
Why mixed-category prioritization fails
The problem is not that they lack ideas; it is that they try to prioritize fundamentally different kinds of systems as if they were the same thing.
Imagine prioritizing "build a simple email auto-responder" against "build a multi-agent network that coordinates sales and marketing operations." The first ships in 2 weeks with predictable ROI. The second is a multi-quarter research project with uncertain outcomes. Comparing them on the same scale produces nonsense. You either underestimate the second or overestimate the first.
The fix is sorting agent initiatives into their architectural category before prioritizing. Then prioritize within categories, and allocate capacity across categories deliberately.
Category 1: Deterministic Automation (predefined workflow with LLM nodes)
What it is
A predefined workflow where one or two steps use an LLM to handle text, classification, or generation. The control flow is deterministic; the LLM is a node, not the brain.
Examples
- Email support: incoming ticket → LLM classifies intent → routes to template → human reviews and sends
- Lead enrichment: new lead → LLM extracts company info from public sources → updates CRM
- Content moderation: user post → LLM scores for policy violations → flags for human or auto-rejects
Why to start here
60-70% of agent opportunities fit Category 1. The pattern is well understood, the cost is predictable, and the ROI is measurable within weeks. A real example from the source piece: an email support agent reached 52% task completion in week 1 and 87% by week 8, generating $18K/month in savings. That kind of clear ROI almost never appears in Category 2 or 3 within the same timeframe.
Roadmap implication
Most teams should have 70% of their agent capacity in Category 1 work. It pays for the experiments in Categories 2 and 3.
Category 2: Reasoning and Acting Agents (ReAct)
What it is
The LLM decides what to do next dynamically. It selects from a set of tools, runs them, reads the output, decides the next step. The control flow is not predetermined; the agent reasons about it.
Examples
- Shopping assistant: user asks a question → agent decides whether to query inventory, check prices, suggest alternatives, or escalate to human
- Research agent: user asks a question → agent searches, reads, synthesizes, decides if more search is needed, produces report
- Customer onboarding: user signs up → agent decides which welcome flow to run based on company size, industry, and stated goals
Why it is harder
The agent makes decisions that the team did not pre-specify. This means the agent will sometimes do things the team did not anticipate. Quality assurance is harder. Failure modes are larger. A shopping assistant in production might improve task completion from 71% to 86% over months of iteration, with conversion lift moving from +8% to +22%. The numbers are great but they take months and the variability is real.
Roadmap implication
25-30% of agent capacity. Pick high-value, well-bounded use cases. The boundary matters more than the ambition: an agent for a clearly defined task is far more shippable than an agent for an open-ended one.
Category 3: Multi-Agent Networks
What it is
Multiple specialized agents coordinating across domains, with one orchestrator agent or peer-to-peer messaging between them. The system has emergent behavior; even the team that built it cannot fully predict what it will do in every situation.
Examples
- A sales pipeline with one agent for outreach, one for qualification, one for proposal generation, all coordinating
- A code-development pipeline with one agent planning, one coding, one testing, one reviewing
- A customer service ecosystem with specialized agents for support, billing, retention, and escalation
Why it is the riskiest category
Multi-agent systems have all the challenges of Category 2 multiplied by the number of agents, plus emergent coordination problems. Debugging is hard. Reliability is hard. The state of the art is still moving fast, which means today's best practices will look naive in a year.
Roadmap implication
5-10% of capacity, mostly as exploration or research bets. The expected outcome is learning, not shipping a stable production system. Treat it like R&D budget.
The allocation pattern that works
| Category | % of agent capacity | Expected ROI timeline | Risk profile |
|---|---|---|---|
| 1: Deterministic | 60-70% | Weeks | Low |
| 2: ReAct | 25-30% | Months | Medium |
| 3: Multi-Agent | 5-10% | Quarters | High |
This allocation produces a portfolio: most of the budget delivers near-term value, a meaningful chunk explores the next frontier, and a small bet probes the horizon. Teams that put 100% in Category 1 stagnate. Teams that put 50% in Category 3 burn out.
How to prioritize within Category 1
Inside Category 1 specifically, the prioritization is easier because the math is more predictable. Score each candidate on:
- Volume: how many times per day or week does this workflow run?
- Cost-per-run before automation: human minutes or system cost
- Confidence in LLM accuracy: how well does the LLM handle this type of task today?
- Failure cost: what happens if the LLM gets it wrong?
High volume, high cost-per-run, high LLM confidence, low failure cost = top of the list. The email support agent example scores well on all four: thousands of tickets, several minutes each, LLMs handle classification reliably, and the worst case is a human reviews before sending.
The common roadmap mistake
Teams skip the categorization step and end up with mixed-category roadmaps where the small Category 1 wins get deprioritized for the visible Category 3 ambitions that never ship. Twelve months later, the team has zero ROI from agents and a Category 3 prototype that still does not work reliably.
Categorize first. Prioritize within categories. Allocate across categories deliberately. The roadmap that emerges is shippable.
The takeaway
Not all AI agents are the same architecturally, and treating them as one category makes prioritization impossible. The three categories (deterministic automation, ReAct agents, multi-agent networks) have very different costs, timelines, and risk profiles. The allocation that works: 60-70% in Category 1 for near-term ROI, 25-30% in Category 2 for exploration, 5-10% in Category 3 as research. Within categories, prioritize on volume, cost, LLM confidence, and failure cost. The teams that ship agent value in 2026 are the ones that started with the right category and resisted the temptation to lead with the most ambitious one.
