Drop your ASIN - get an optimized listing and push it to Amazon Seller Central in one click. Drop your ASIN - get an optimized listing, push to Amazon. Try it →Academic AI taxonomy (reactive, model-based, goal-based, utility-based, learning) is technically correct and almost useless when you're buying or building. The practical taxonomy that matters in 2026 is by how the agent is triggered, what it's scoped to do, and how much autonomy it has. The classic five-class taxonomy is documented well by IBM Think on AI agents, and Anthropic's Building Effective Agents covers the production-side distinctions. This page covers seven types that actually map to product decisions.
Group AI agents by three dimensions: trigger (human-activated vs event-activated), scope (task-specific vs generic), and autonomy (assistive vs autonomous). The seven combinations that actually show up in production are below.
Forget the Russell-and-Norvig textbook classification for a moment. In 2026 ecommerce, what determines whether an agent fits a job is:
Combine those three dimensions and you get the agent types you'll actually encounter. Below are the seven that matter most.
The most common type for ecommerce sellers in 2026. You click "Optimize this listing." The agent runs. It returns a draft. You review and approve.
Trigger is a human click. Scope is narrow (one job: listing optimization). Autonomy is assistive (drafts for human approval, doesn't auto-apply).
Examples: most agents on the SellerShorts marketplace, Helium 10 Listing Builder runs, Jungle Scout AI Assist sessions.
Best for: high-value or high-risk work where human judgment is non-negotiable. Listing copy that ships to live Amazon pages. Image generation for a brand launch. A+ content drafts.
Same as Type 1 but the agent executes the output without a review step. You click, the agent runs, the action happens. Examples are repricing tools that adjust prices in real-time and PPC tools that update bids on a schedule.
Trigger is human (you set the rules). Scope is narrow. Autonomy is high (within the rules).
Best for: bounded operational work where the consequences of any single action are small and reversible. Bid changes within a tight ACoS range. Price changes within a tight margin range.
Risky for: anything where a single wrong action is expensive. Auto-posted reviews. Auto-approved buyer refunds. Auto-deleted listings.
Something happens in the world (a new negative review, a stockout warning, a competitor price change). The agent runs automatically. It does the work and notifies you to review.
Trigger is an event (a webhook, a polling cycle, a threshold breach). Scope is narrow. Autonomy is assistive (it acts but waits for you to sign off).
Examples: review monitoring agents that draft responses on new negatives, inventory forecasters that flag stockouts before they happen, brand-monitoring agents that catch hijacked listings.
Best for: ongoing operational monitoring where speed of detection matters but final action still needs a human in the loop.
Like Type 3, but the agent executes without review. An event fires, the agent runs, it acts on the world. This is the type with the highest scaling potential and the highest risk.
Trigger is event. Scope is narrow. Autonomy is high.
Examples: continuous repricers, dynamic bid managers, auto-replenishment systems.
Best for: tightly bounded operational work that's been validated. The pre-conditions matter: you need rules, guardrails, kill switches, observability. The Amazon BSA Agent Policy that took effect March 4, 2026 added explicit constraints on what Type 4 agents can do on a seller's account. Worth reading before deploying one.
You ask a general-purpose AI assistant a question. It works on a wide range of tasks. ChatGPT, Claude, Amazon's Seller Assistant. Not bounded to one job.
Trigger is human (you typing). Scope is broad. Autonomy is assistive (turn by turn).
Best for: ad-hoc work that doesn't fit a pre-built specialized agent. Drafting an unusual product description. Brainstorming a launch plan. Researching a competitor's brand.
Limitation: jack of all trades, master of none. Specialized agents typically beat generic ones on specific repeatable tasks.
Several agents working together on a task too complex for one. Covered in depth in the multi-agent systems guide. The shape is: an orchestrator agent decomposes the job, dispatches sub-tasks to specialized worker agents, and stitches results together.
Trigger can be either. Scope is broad (covers the whole orchestrated job). Autonomy varies (often high within the system, with human review at the end).
Best for: complex workflows where one agent can't hold the whole picture. A full quarterly Amazon audit, a brand-launch creative package, a cross-channel campaign rollout.
A growing category in 2026. AI assistants embedded inside a specific tool, with deep knowledge of that tool's data. Amazon's Seller Assistant is the canonical example for Amazon sellers. GitHub Copilot Workspace is the developer equivalent.
Trigger is human (you're inside the tool, you ask the copilot). Scope is broad within the tool's domain. Autonomy is mostly assistive, but some copilots can execute actions inside the host application.
Best for: getting more out of a tool you already use. Faster onboarding. Smarter UI navigation. Account-context answers.
This dimension matters enough to deserve its own section, since the deleted /human-activated-vs-event-activated page is now absorbed here.
Run when a person decides to run them. The seller clicks a button, types into a prompt, uploads a file. The agent executes once, returns a result, stops.
Pros: explicit user intent, low risk of unexpected runs, easy to budget (you control the spend).
Cons: capped by the human's attention. If you forget to run the agent, the work doesn't get done. Scaling is bottlenecked by how often you remember.
Run automatically when a condition is met. A new review arrives. Inventory drops below 30 days. ACoS exceeds the target by 20%. The agent fires without anyone clicking.
Pros: continuous coverage, faster response, scales beyond human attention.
Cons: higher cost variability (runs depend on event volume), harder to reason about, easier to cause unintended damage if guardrails are missing.
Quick rule: human-activated when the cost of skipping a run is low and the cost of a wrong run is high. Event-activated when the cost of skipping a run is high and the cost of a wrong run is low (or bounded).
Listing optimization: human-activated (skipping is fine, wrong run is high-impact). Inventory forecasting: event-activated (skipping = stockout, wrong run = a redundant email).
Another dimension that deserves its own section, since the deleted /task-specific-vs-generic-tools page is now here.
Built to do one job well. Listing optimizer. Image generator. PPC bid manager. The system prompt, tools, and memory are all tuned for that one job.
Pros: higher reliability on the target job, more predictable cost, easier to evaluate.
Cons: doesn't help with anything outside its scope. You need a different agent for each different job.
Built to do many jobs reasonably well. ChatGPT, Claude.ai, Amazon Seller Assistant. The same agent handles whatever you ask.
Pros: covers wide territory, lower cognitive overhead (one tool instead of many).
Cons: weaker on any individual specialized job. Generic agents trying to do listing optimization will produce mediocre listings compared to a dedicated listing optimizer.
Generic agent for ad-hoc thinking work and one-off questions. Task-specific agents for the repeated operational jobs that justify a dedicated tool. Most serious Amazon sellers in 2026 run this hybrid pattern: one assistant for the variable work, a stack of task-specific agents for the recurring work.
| Type | Trigger | Scope | Autonomy | Example |
|---|---|---|---|---|
| 1 | Human | Task-specific | Assistive | SellerShorts listing-optimizer run |
| 2 | Human | Task-specific | Autonomous | Bounded repricer |
| 3 | Event | Task-specific | Assistive | Review-monitor with draft responses |
| 4 | Event | Task-specific | Autonomous | Continuous bid manager |
| 5 | Human | Generic | Assistive | ChatGPT, Amazon Seller Assistant |
| 6 | Either | Broad (orchestrated) | Varies | Multi-agent quarterly audit |
| 7 | Human | Domain-specific | Mostly assistive | Amazon Seller Assistant, GitHub Copilot |
For completeness, the Russell-and-Norvig classification still gets cited in serious AI literature. Brief glossary:
These are useful for university courses. They're rarely useful for product decisions because most production agents are mixtures (goal-based + learning + utility, etc.) and the categories don't map cleanly to "should I buy this." Use the practical taxonomy above for that.
The practical taxonomy that matters in production groups agents by trigger (human vs event-activated), scope (task-specific vs generic), and autonomy (assistive vs autonomous). The seven combinations that actually show up in ecommerce are: human-task-assistive, human-task-autonomous, event-task-assistive, event-task-autonomous, human-generic-assistive, multi-agent orchestrated, and domain-specific copilots like Amazon Seller Assistant.
Assistive agents produce a draft and wait for human review before anything ships. Autonomous agents execute directly. Assistive is right for high-value or high-risk work like listing copy and image generation. Autonomous is appropriate for tightly bounded operational work like bid changes inside a tight ACoS range, but only with guardrails and a kill switch.
Use a generic assistant (ChatGPT, Claude, Amazon Seller Assistant) for ad-hoc work that does not fit a pre-built tool. Use task-specific agents for repeated operational jobs like listing optimization or PPC analysis. Most serious Amazon sellers in 2026 run a hybrid: one assistant for variable work, a stack of task-specific agents for recurring work.
Yes, but with constraints. Amazon's BSA Agent Policy took effect March 4, 2026 and added explicit rules on what autonomous agents can do on a seller's account. Continuous repricers and dynamic bid managers are still allowed but the vendor must comply with the policy's identification, scope, and audit requirements. Always ask the vendor whether they meet the BSA Agent Policy before turning on autonomous execution.
A multi-agent system has several specialized agents working together on a task too complex for one. An orchestrator agent decomposes the job, dispatches sub-tasks to worker agents, and stitches results together. Good fit for quarterly Amazon audits, brand-launch creative packages, and cross-channel campaign rollouts.
SellerShorts focuses on Type 1 agents (human-activated, task-specific, assistive). Drop in your inputs, get drafts for review, approve before anything ships to your Amazon account.
Browse SellerShorts agents
sellershorts