Workflow-Level Control
How much of the end-to-end workflow Zevo runs.
§ 01 — Overview
zevo describe --overview
Objective in, evolved model out
Zevo handles everything in between.
Scroll the figure sideways →
“Could you build me a model that excels at medical reasoning?”
User ObjectiveEvolved ModelA medicine-specialized evolved model
Zevo
self-improvement
Baseline
Proposebetter training
Evaluate
Keep the best
Three properties of Zevo
Autonomous
- End-to-end multi-agent collaboration
- Self-improves toward your objective
Reliability
- Scored only by the evaluation harness you set
- The best-scoring model is what you get
Self-hosted
- Deploys on your own devices
- Agents can run fully sandboxed
§ 02 — Architecture
zevo status --architecture
Seven agents, one loop
Seven agents on one line to run the evolving loop.
Scroll the figure sideways →
Zevo iteration
Input
User Objective
Output
Evolved Model
Click any agent to read what it does specifically.
§ 03 — Demo
zevo demo --replay
System in action
§ 04 — Interface
zevo drive --interface
Two ways to drive it
Zevo runs from CLI or UI. Same run, either way.
CLI
Run it from the terminal
UI
Run it from the browser
01 / 08
Dashboard
§ 05 — Control
zevo describe --control-levels
Two dimensions of control
Workflow and Optimization control.
Optimization-Level Control
How much of the optimization setup Zevo decides.
Workflow-Level Control
| Mode | End-to-End Workflow | Zevo Creates Evaluation | Unconstrained Agents |
|---|---|---|---|
| M1Single StageThe user sends one focused task to one specialized agent. | ✕ | ✕ | ✕ |
| M2CustomizedThe user provides the evaluation and adds constraints to selected agents. Zevo runs the complete workflow. | ✓ | ✕ | ✕ |
| M3StandardThe user provides the evaluation. Zevo runs the complete workflow. | ✓ | ✕ | ✓ |
| M4AutoZevo creates the evaluation and runs the complete workflow. | ✓ | ✓ | ✓ |
M1
Single Stage
The user sends one focused task to one specialized agent.
- End-to-End Workflow
- ✕
- Zevo Creates Evaluation
- ✕
- Unconstrained Agents
- ✕
M2
Customized
The user provides the evaluation and adds constraints to selected agents. Zevo runs the complete workflow.
- End-to-End Workflow
- ✓
- Zevo Creates Evaluation
- ✕
- Unconstrained Agents
- ✕
M3
Standard
The user provides the evaluation. Zevo runs the complete workflow.
- End-to-End Workflow
- ✓
- Zevo Creates Evaluation
- ✕
- Unconstrained Agents
- ✓
M4
Auto
Zevo creates the evaluation and runs the complete workflow.
- End-to-End Workflow
- ✓
- Zevo Creates Evaluation
- ✓
- Unconstrained Agents
- ✓
Optimization-Level Control
| Level | Training Data | Training Model | Training Method |
|---|---|---|---|
| L1Entry AutonomousZevo iterates within the given data, model, and method. | User | User | User |
| L2Constraint AutonomousZevo collects and even creates the training data. | Zevo | User | User |
| L3Partially AutonomousZevo picks the training method as well. | Zevo | User | Zevo |
| L4Fully AutonomousZevo decides everything from the objective alone. | Zevo | Zevo | Zevo |
L1
Entry Autonomous
Zevo iterates within the given data, model, and method.
- Training Data
- User
- Training Model
- User
- Training Method
- User
L2
Constraint Autonomous
Zevo collects and even creates the training data.
- Training Data
- Zevo
- Training Model
- User
- Training Method
- User
L3
Partially Autonomous
Zevo picks the training method as well.
- Training Data
- Zevo
- Training Model
- User
- Training Method
- Zevo
L4
Fully Autonomous
Zevo decides everything from the objective alone.
- Training Data
- Zevo
- Training Model
- Zevo
- Training Method
- Zevo
Result
zevo report --run
Example Run
L1 task: Turn the base model into an instruction-following model
- Training data
- trl-lib/Capybara (train set)
- Training model
- OLMo-2-0425-1B
- Training method
- Full SFT
Scroll the figure sideways →
Iteration
0
1
2
3
4
5
6
7
How Zevo evolves the model
Baseline
Full SFT, LR 2e‑5, 2 epochs
Epochs 2 → 3
LR 2e‑5 → 1e‑5
Dropped 6.5% of the data
LR 2e‑5 → 3e‑5
LR 3e‑5 → 5e‑5
LR 5e‑5 → 8e‑5
How Zevo self-improves
The bar to beat
Val loss flat, not up → epochs have headroom
Extra epoch bought nothing → try the LR
Val loss still falling → underfit, not overfit
Data is not the wall → take the LR up
Plateau broken → peak not found yet
Still rising → keep going up
Zevo stopped the run here
Waitlist
zevo waitlist --join
Join the waitlist
We are evaluating Zevo and deploying the online system. Join the waitlist and stay tuned.