Tom: The Cognitive Worker That Never Says No
Zalo knows, Lisa orchestrates, Hermes executes. But someone has to process the raw material. Tom classifies, extracts, summarizes and transforms so the others can work. 4 model tiers, automatic routing, 260ms average. It's the bridge between all agents.
Gonzalo Monzón
July 7, 2026 · Series: Agentes Cadences — The 4 Pillars (3/4)
TL;DR
Tom is the cognitive worker of the ecosystem. 4 model tiers (FAST, CHEAP, GRANITE, QWEN) with automatic routing — each task gets the right model without manual configuration. It lives on Cloudflare Workers, has 7 MCP tools, and is the only agent that everyone else calls. Its superpower: it never blocks. Its weakness: it doesn't distinguish context nuances. This is the third of 4 articles about the Cadences Lab agents.
What is Tom?
In the Cadences ecosystem, agents have very defined roles. Zalo knows, decides and coordinates. Lisa analyzes, plans and evaluates. Hermes builds, deploys and executes.
But all of this requires a prior step: understanding the world. Messages arrive raw. Texts come without structure. Requests are ambiguous. Someone has to classify, extract, summarize and transform before the others can work.
That someone is Tom.
Personality: Tom is 19 years old — metaphorically. The youngest of the team. Its tone is relaxed, direct, sometimes irreverent. The colleague who tells you "dude, this is simpler than it looks." Before processing something it says "feed me something" — its way of saying "prep the data for me."
This personality isn't a whim. In a multi-agent system, each agent needs a distinct voice so logs are readable, debugging is intuitive, and the system feels like a real team.
The Only One That Never Blocks
Tom never blocks. Zalo hesitates when it lacks context. Lisa hesitates when there's no prior pattern. Hermes hesitates when execution is ambiguous. Tom simply executes. There's no task it refuses — it just asks "which model do we use?" and starts.
It's not bravery. It's design. Tom doesn't make decisions or evaluate consequences. Its job is to process and deliver. There's no possibility of blocking because there's no decision to make.
The Bridge Between Agents
Tom is the only agent that everyone else calls. It's the glue that connects the ecosystem:
| Who calls | For what | What Tom does |
|---|---|---|
| Zalo | "Classify this message" | Picks Granite → 0.017¢, 95% accuracy |
| Lisa | "Extract data from this text" | Picks Qwen → structured JSON |
| Hermes | "Summarize this before acting" | Picks FAST → 330ms, good enough |
| Zalo | "Is this text urgent?" | Classifies with CHEAP → 260ms |
Real case: Zalo receives a message: "I got the invoice but the total doesn't add up." Zalo needs to know if it's urgent and what type it is. Instead of doing it itself (would take 2s and consume valuable context), it calls Tom. Tom responds in 260ms with CHEAP: "urgent: yes, category: billing." Zalo can decide without distraction. Tom processes and delivers. It passes the ball.
The Weakness: No Context Nuance
Tom doesn't distinguish contextual nuances. For it, an angry customer email and a clinical report are both "text to classify." It doesn't understand architecture, can't weigh technical severity — that's Zalo's job. That's why it never acts alone. Tom processes, but doesn't decide.
4 Model Tiers with Intelligent Routing
Tom has 4 models. But the magic isn't that it has them — it's that it automatically chooses which one to use based on the task, without manual configuration:
| Tier | Model | Cost | Latency | When it's used |
|---|---|---|---|---|
| ⚡ FAST | Llama 3.1 8B FP8 | 0.045¢ | ~150ms | Summaries, transformations, direct responses |
| 💰 CHEAP | Llama 3.2 3B | 0.051¢ | ~400ms | Simple classifications, basic extraction |
| 📉 GRANITE | IBM Granite 4.0 Micro | 0.017¢ | ~1.2s | Precise classification when cost matters |
| 🔬 QWEN | Qwen3 30B A3B | 0.051¢ | ~2.5s | Complex extraction, text reasoning |
Routing criteria:
Urgent? → FAST (~150ms), fallback → CHEAP (~400ms)
Simple classification? → CHEAP (~400ms, 0.051¢)
Cost-critical? → GRANITE (0.017¢, ~1.2s)
Needs understanding? → QWEN (~2.5s, complex extraction)
No manual configuration. Tom evaluates the request and chooses in milliseconds. The result is that each task pays exactly the cost it needs — neither more expensive nor slower than necessary.
7 MCP Tools for Every Type of Process
Tom exposes 7 tools through MCP, each designed for a type of task:
📋 tom_process
Processes text with a free prompt. Uses FAST for speed or GRANITE for simple tasks
🔤 tom_classify
Classifies text into predefined categories using GRANITE (ultra cheap)
📄 tom_extract
Extracts structured data from text as JSON using QWEN
📝 tom_summarize
Summarizes long content with GRANITE for extensive texts
🔄 tom_transform
Transforms content between formats (JSON, CSV, YAML, HTML)
📦 tom_batch
Processes multiple texts with the same instruction, in batch
❤️ tom_status
Tom's status and stats: usage, accumulated costs, average latency
Edge Native, Like Its Siblings
Tom lives on Cloudflare Workers with service bindings to Zalo, Lisa and PDB. But Tom doesn't write code, doesn't deploy, has no filesystem access.
Tom classifies, extracts, summarizes and transforms. Nothing else. And that's enough to be the glue of the ecosystem.
The 4 Agents, One Ecosystem
| Dimension | Zalo | Lisa | Tom | Hermes |
|---|---|---|---|---|
| Role | Knows and coordinates | Analyzes and evaluates | Processes and delivers | Executes and deploys |
| Strength | Living mental map | Incoherency detection | Never blocks | Executes anything |
| Weakness | Doesn't execute code | Slow without patterns | No context nuance | No deep reasoning |
| Model | DeepSeek V4 | Llama 3.2 + DeepSeek Pro | 4 tiers (FAST→QWEN) | Session model |
The Youngest. The One That Processes the Most.
Tom is the youngest on the team. And the one that processes the most. Its intelligent routing across 4 model tiers ensures each task gets the right balance of speed, cost and quality. No manual configuration. No decisions to make.
It doesn't grasp nuance. It doesn't understand politics. But when you need to classify, extract, summarize or transform — fast, cheap and well — Tom is who you call. And it never says no.
Series: Agentes Cadences — The 4 Pillars (3/4)
← Previous
Lisa — The OrchestratorNext →
Hermes — The Execution LayerTom is the youngest. And the one that processes the most.
Hermes closes the circle: executing, deploying, and coordinating the physical world. The last of the series.
Tom is property of Cadences Lab. Agents Series — 4 articles. © 2026.