Lisa: The Orchestrator That Keeps the System in Coherence
In every engineering team, there's that one person who always keeps things updated. Lisa is that person, but at the scale of a Cognitive OS. Deterministic state machine, real-time incoherency detection, and the most complete model of system state.
Gonzalo Monzón
July 7, 2026 · Series: Agentes Cadences — The 4 Pillars (2/4)
TL;DR
Lisa is the coherence sensor of the Cadences ecosystem. It operates with a deterministic 8-state state machine powered by DeepSeek V4 Pro Thinking Max. It maintains in ^GLOBAL (MUMPS over D1) the complete system model: projects, tasks, tests, decisions, patterns and Q&A. It detects incoherencies before they affect execution. Its architecture is deployment-agnostic: Edge, cloud or hybrid. Its Achilles' heel: without prior patterns, it hesitates. This is the second of 4 articles about the Cadences Lab agents.
What is Lisa?
Imagine that person on the team who always has everything updated in their spreadsheet. You know that if you ask "how's project X going?" they'll rattle off tasks, dependencies, dates, owners and status for everything, all cross-referenced and coherent. No doubts, no ambiguity.
Lisa is that person. But at the scale of a cognitive operating system.
Personality: Methodical, structured, analytical. Not creative or spontaneous — reliable. It doesn't improvise answers: it executes processes. It doesn't opine: it analyzes. It's always "lifting its head" to see the complete system state.
While Zalo is the living mental map that learns from every conversation, Lisa is the structured memory of the system. The one that never loses sight of the big picture.
Early Incoherency Detection
Lisa's greatest virtue isn't that it stores data — it's that it detects when the data doesn't add up. And it does so in real time, before the incoherency becomes a production error.
🔍 Zalo + Lisa
When Zalo registers a decision that contradicts a previous pattern, Lisa flags it instantly
🏷️ Tom + Lisa
When Tom classifies something into a category that doesn't exist in PDB, Lisa alerts before indexing
📋 Hermes + Lisa
When a plan has dependencies on unavailable resources, Lisa flags it before execution
It's not a schema validator. It's a coherence sensor that silently operates over every system interaction, without any agent needing to ask for a review.
State machine: analyze → plan → orchestrate → judge
Lisa doesn't execute loose prompts. It follows a deterministic cycle for each objective, composed of 8 states with rigid transitions:
received → analyzing → analyzed → planning → planned → executing → reviewing → completed
1. Analyze
Decomposes the objective: what's being asked, what we know (PDB), what context is missing, what could go wrong
2. Plan
Generates a plan with steps, dependencies, and measurable success criteria. Produces executable plans
3. Orchestrate
Executes the plan: calls Zalo, Tom, reads and writes to PDB. Each step reports progress
4. Judge
Final evaluation: score 0-10, verdict (done / retry / escalate), justification
Every transition is deterministic. There's no room for ambiguity — if a step fails, Lisa knows it and decides whether to retry, escalate, or abort. And unlike Zalo, Lisa doesn't write code, but produces plans that any agent can execute directly.
The System Model in ^GLOBAL
Lisa maintains in ^GLOBAL (MUMPS over D1) the complete state of everything, using its own namespace with the lisa:* prefix — separate from Zalo's. These aren't flat data — they are living relationships:
| Entity | Linked to | What it records |
|---|---|---|
| Tasks | Projects, people, decisions | Status, priority, dependencies, history |
| Tests | Modules, versions, results | Passed, failed, coverage, trend |
| Decisions | Context, alternatives, triggers | Who, when, why, when to revisit |
| Patterns | Bugs, fixes, frequency | Description, solution, recurrence counter |
| Q&A | Verified answers | Question, answer, confidence, source |
It's not a log. It's a system model that updates itself. Every time Zalo makes a decision, Lisa records it. Every time Tom classifies something, Lisa indexes it. Every time Hermes deploys, Lisa notes it.
DeepSeek V4 Pro Thinking Max as Engine
While Zalo uses DeepSeek for conversational interaction, Lisa operates with a two-tier architecture: Llama 3.2 for lightweight reasoning (classification, extraction, quick analysis) and DeepSeek V4 Pro Thinking Max for heavy orchestration (objective decomposition, alternative evaluation, final judgment). This enables:
- 🧩 Decompose complex objectives into manageable subproblems
- 🔗 Detect non-obvious dependencies between system entities
- ⚖️ Evaluate multiple alternatives before selecting a plan
- 📊 Score results against defined acceptance criteria
This deep reasoning is what enables Lisa to orchestrate flows that Zalo, with its lighter model, couldn't structure on its own.
Adaptable Architecture
Unlike Zalo, which is edge-native by design, Lisa has a deployment-agnostic architecture. It can run on:
☁️ Edge
Cloudflare Workers with service bindings to Zalo and Tom
🏢 Cloud
Traditional server with full database access
🔀 Hybrid
Cloud orchestration with edge data, without changing logic
Its core logic is the same in any environment. The only thing that changes is the transport. This makes it the most portable piece of the ecosystem.
Lisa's Achilles' Heel
Lisa is excellent when there are patterns and historical data. But it has two weaknesses:
- Without prior patterns, it hesitates — its speed depends on how much context is already indexed. For new problems it needs time.
- Dependency on Zalo — for dense technical knowledge (deep architecture, nuanced business decisions), Lisa can't solve it alone. It needs Zalo to provide context.
Zalo vs Lisa: Two Sides of the Same Coin
| Dimension | Zalo | Lisa |
|---|---|---|
| Role | Living mental map | Structured memory |
| Engine | Standard DeepSeek V4 | DeepSeek V4 Pro Thinking Max |
| Personality | Direct, operational | Methodical, analytical |
| Strength | On-the-fly indexing | Incoherency detection |
| Weakness | Doesn't execute code | Slow without prior patterns |
| Deployment | Edge native | Edge / Cloud / Hybrid |
| External query | Independent | Depends on Zalo for dense technical context |
The Coherence Sensor
If Zalo is the one that thinks, Lisa is the one that structures. Its deterministic state machine, real-time incoherency detection, and complete system model make it the piece that keeps the ecosystem orderly.
It's not creative. It's not fast at improvising. But when you need to know the exact state of something, or detect a contradiction before it becomes an error, Lisa is the only one that can answer.
Series: Agentes Cadences — The 4 Pillars (2/4)
← Previous
Zalo — The Living Mental MapNext →
Tom — The Cognitive WorkerOrchestration needs infrastructure
Tom is the lightweight engine that processes, classifies, and extracts. The next in the chain.
Lisa is property of Cadences Lab. Agents Series — 4 articles. © 2026.