LotsAgentWhat an AI Agent Actually Costs to Run: A Real Breakdown for Lean Teams
You want an AI agent that does real work — triages support tickets, drafts emails, qualifies leads, runs across channels. The question isn't whether it's possible. It's whether the numbers add up.
Most cost comparisons stop at the model API bill: a few dollars per million tokens. That's like saying a car costs what the tires cost. It misses everything that makes the car drivable.
Here's the full picture — one-time build costs, running costs, and the maintenance costs that quietly accumulate — so you can decide what makes sense for your team.
The Three Cost Layers of Running an AI Agent
Before comparing platforms, it helps to separate the costs into three distinct layers:
Layer 1: One-time build cost — The upfront engineering to go from zero to a working agent. Infrastructure setup, tool wiring, authentication flows, deployment configuration.
Layer 2: Running cost — What it costs to actually run the agent day-to-day. Execution compute, model API calls, storage, channel delivery.
Layer 3: Maintenance cost — What breaks over time. OAuth tokens that expire. API rate limits that need retry logic. Schema changes that break tool connections. Audit trails that need storage. Model updates that change behavior.
Most teams budget for Layer 1. Some budget for Layer 2. Almost nobody budgets correctly for Layer 3 — and that's where the real difference between approaches shows up.
Path 1: Building In-House (DIY Infrastructure)
If you're technical, building an agent from scratch looks appealing. You control everything. No platform lock-in. Full flexibility.
Here's what actually goes into it:
Memory Infrastructure
Your agent needs to remember context across conversations — user-specific memory, agent-specific memory, and the ability to search past interactions. That means running or paying for a vector store.
- Pinecone: Starts at $70/month for the entry tier that supports production workloads
- Weaviate Cloud: Similar range, $50–100/month depending on vector dimensions and throughput
- Self-hosting Qdrant or Milvus: $0 software, but you pay for the compute and ops time
Tool Integrations (OAuth + API Wiring)
Each tool your agent connects to needs OAuth flows, token refresh handling, scoped permissions, and error handling. A support agent using Gmail, Slack, and a CRM needs three separate OAuth implementations — each with its own refresh token lifecycle, rate limit handling, and error states.
Engineering estimate: 2–5 days per integration depending on the provider's API quality.
Durable Execution and Retry Logic
Production workflows fail. A network timeout at step 3 of 6 means either the whole workflow restarts from scratch (wasteful) or you build checkpoint logic (expensive).
Building durable execution from scratch requires:
- A queue system (Redis, SQS, or similar)
- Checkpoint persistence per step
- Retry policies with exponential backoff
- Dead-letter handling for unrecoverable failures
This is non-trivial infrastructure. Most teams underestimate it by at least 2–3 engineering weeks.
Multi-Channel Deployment
Running the same agent on web chat and email and Telegram and API means building or integrating four separate channel adapters — each with different message formats, auth models, and delivery guarantees.
Audit Logging and Identity
If your agent acts on behalf of your business, you need to know what it did, when, and under whose authority. Building an audit trail that captures every action, tool call, and decision point means a logging pipeline, storage, and a queryable interface.
Estimated DIY Bill for a Production Agent
| Component | Setup Time | Ongoing Monthly Cost |
|---|---|---|
| Vector memory store | 2–5 days | $50–100 |
| Tool integration wiring | 5–10 days | — |
| Durable execution infra | 10–15 days | $50–200 |
| Multi-channel setup | 5–10 days | — |
| Audit pipeline | 3–5 days | $20–50 |
| Ongoing maintenance | 4–8 hrs/week | — |
Estimated first-month total: $120–350 in services + 3–6 engineering weeks.
Path 2: DIY with Agent Frameworks (LangChain, n8n, etc.)
Frameworks reduce the one-time build cost. You get abstractions for memory, tools, and chains out of the box.
What they don't include:
- Hosted infrastructure — You still provision servers, databases, and storage
- Multi-channel deployment — The framework gives you building blocks, not deployed channels
- Production retry logic — Framework primitives for chains, but no durable execution with checkpointing
- Managed OAuth — Token refresh and lifecycle management are still your problem
- Ongoing updates — Framework version changes can break your agent config
The framework path lowers the upfront build from 6 weeks to roughly 2–3 weeks, but most ongoing operational costs remain the same.
Path 3: AI Agent Platform (LotsAgent)
An AI agent platform like LotsAgent includes all three cost layers in one setup. You don't build infrastructure; you configure the agent.
What's Included
- Persistent memory — Text and vector memory included. User-specific and agent-specific. No separate database to provision.
- 100+ tool integrations — Pre-connected via Composio. OAuth flows, token refresh, rate limit handling included.
- Durable execution — Powered by Inngest. Every execution checkpointed. Failures retry from the last successful step, not from scratch.
- Multi-channel deployment — Configure once. Available via web UI, email, Telegram, API, MCP, and webhooks from the same setup.
- Full audit trail — Every action logged. Every execution visible. Irreversible actions configurable for human approval.
- Skills system — Build a workflow once, package it as a skill, reuse across agents.
What It Costs
| Plan | Price | What You Get |
|---|---|---|
| Free | $0 | 10 agents, 1,000 executions/month, text memory |
| Pro | $20/month | Unlimited agents, 10,000 executions included, vector memory, $0.01/additional execution |
Model API costs are separate — you pay for the model tokens whether you use a platform or DIY. LotsAgent supports OpenRouter, OpenAI, Anthropic, and BYOK so you control exactly which model and provider you use.
Side-by-Side Cost Comparison
Workload A: Small — Single Channel, Low Volume
Support agent, one channel (web chat), ~500 conversations/month
| Cost Item | DIY | Framework (n8n/LangChain) | LotsAgent |
|---|---|---|---|
| Infrastructure setup | $2,500–7,500 (2–3 wks) | $1,500–3,500 (1–2 wks) | $0 |
| Monthly infra/services | $100–250 | $100–250 | $0 (Free) or $20 (Pro) |
| Model API | $10–30 | $10–30 | $10–30 |
| Maintenance/month | $400–800 (4–8 hrs) | $200–500 (2–5 hrs) | $0 |
| Year 1 Total | $9,000–19,000 | $5,500–11,000 | $120–600 |
Workload B: Medium — Multi-Channel, Moderate Volume
Sales triage agent, 3 channels (web, email, Telegram), ~3,000 interactions/month
| Cost Item | DIY | Framework | LotsAgent |
|---|---|---|---|
| Infrastructure setup | $7,500–15,000 (4–6 wks) | $3,500–7,500 (2–3 wks) | $0 |
| Monthly infra/services | $250–500 | $250–400 | $20 (Pro) |
| Additional executions | — | — | $0 (within 10K) |
| Model API | $50–150 | $50–150 | $50–150 |
| Maintenance/month | $800–1,500 (8–15 hrs) | $500–800 (5–8 hrs) | $0 |
| Year 1 Total | $20,000–40,000 | $12,000–24,000 | $840–2,040 |
Workload C: Large — Multi-Agent, High Volume
Full operations: support + sales + triage agents, 5+ channels, ~20,000 executions/month
| Cost Item | DIY | Framework | LotsAgent |
|---|---|---|---|
| Infrastructure setup | $15,000–35,000 (6–8 wks) | $7,500–15,000 (3–5 wks) | $0 |
| Monthly infra/services | $500–1,500 | $400–1,000 | $20 (Pro) + $100 (10K extra execs) |
| Model API | $200–500 | $200–500 | $200–500 |
| Maintenance/month | $1,500–3,000 (15–30 hrs) | $800–1,500 (8–15 hrs) | $0 |
| Year 1 Total | $40,000–90,000 | $22,000–45,000 | $3,840–7,440 |
The Model Cost Question
Model API costs get the most attention but are the most consistent variable across all approaches. Whether you build your own agent or use a platform, you're calling a language model — and you pay the same token price.
LotsAgent doesn't add margins on model tokens. You can use OpenRouter for broad model access, connect directly to OpenAI or Anthropic, or bring your own API key for direct cost control. The platform charge covers everything around the model — memory, tools, channels, durable execution, audit trail.
To estimate your monthly model spend:
Total = (avg tokens per execution × executions per month) × (price per 1K tokens)
For a typical support agent handling 3,000 conversations a month with ~2,000 input + ~500 output tokens each:
- ~7.5M input tokens @ $0.15/1M (Claude 3.5 Haiku) = $1.13
- ~1.5M output tokens @ $0.60/1M = $0.90
- Total model cost: ~$2/month
Even with larger models (GPT-4o, Claude Sonnet), model costs rarely exceed $50–150/month for moderate volumes — a fraction of the infrastructure costs.
When Each Path Makes Sense
Build in-house when: Infrastructure is your product. You're a platform company shipping agent infrastructure to others, or you have specialized requirements no off-the-shelf platform can meet — and you have the engineering capacity to maintain it.
Use a framework when: You need more control than a platform offers but want to skip some boilerplate. Expect to handle your own deployment, scaling, and operations.
Use an AI agent platform like LotsAgent when: The workflow is your product — not the infrastructure. You want agents running in days, not weeks. You want to pay for what you use, not for servers you maintain.
For lean teams, operators, and founders who need agents that do real work without becoming infrastructure engineers, an AI agent platform is the repeatable answer. The math is straightforward: your time is better spent on the workflows the agent handles, not on the scaffolding that runs it.