Evolution Workflow

The evolution workflow improves Gaius agents over time through task ideation, training, evaluation, and promotion. This is a cycle that repeats as agents accumulate more data and experience.

Overview

Status check --> Task ideation --> Training --> Evaluation --> Promotion
     ^                                                           |
     |___________________________________________________________|

Each cycle produces a new generation of agent versions. Successful versions are promoted to active status; underperformers are retained for comparison but not used in production.

Step 1: Check Status

Before starting an evolution cycle, check the current state:

uv run gaius-cli --cmd "/evolve status" --format json

This shows the current generation number, active agents, evaluation state, and any capability gaps. Pay attention to:

Generation: which cycle you are on
Active agents: which agent versions are currently serving
Capability gaps: areas where agents underperform

Step 2: Task Ideation

Generate new training tasks based on identified capability gaps:

uv run gaius-cli --cmd "/evolve task ideation" --format json

The ideation process analyzes recent performance data and gap analysis to propose tasks that target specific weaknesses. Tasks are designed to push agents toward areas where they currently underperform.

Step 3: Trigger Evolution

Start the evolution cycle. This runs training with the generated tasks and produces new agent versions:

uv run gaius-cli --cmd "/evolve trigger" --format json

Evolution requires healthy inference endpoints. Verify with:

uv run gaius-cli --cmd "/gpu status" --format json | \
    jq '.data.endpoints[] | {name, status}'

All endpoints should show HEALTHY before triggering evolution. If they do not, run /health fix endpoints first.

Step 4: Evaluate

After training completes, evaluate the new agent versions against held-out test data:

# Check evaluation results
uv run gaius-cli --cmd "/evolve status" --format json | jq '.data.evaluation'

# View held-out statistics
uv run gaius-cli --cmd "/evolve held-out stats" --format json

Evaluation uses the RASE verification framework. Each agent version is scored on accuracy (0.0-1.0, proportion of constraints satisfied) and compared against previous versions.

Step 5: Promote or Roll Back

If the new version outperforms the current active version, promote it:

# View the best version
uv run gaius-cli --cmd "/evolve best" --format json

# Promote (via MCP or direct command)
uv run gaius-cli --cmd "/evolve promote" --format json

If the new version underperforms, roll back to a known good version:

uv run gaius-cli --cmd "/evolve rollback" --format json

Evolution Daemon

For continuous improvement, start the evolution daemon which runs cycles automatically:

# Start the daemon
uv run gaius-cli --cmd "/evolve daemon start" --format json

# Check daemon status
uv run gaius-cli --cmd "/evolve daemon status" --format json

# Stop the daemon
uv run gaius-cli --cmd "/evolve daemon stop" --format json

The daemon monitors capability gaps and triggers evolution cycles when thresholds are exceeded.

Monitoring Evolution Trends

Track improvement over time:

uv run gaius-cli --cmd "/evolve trend" --format json

This shows how agent performance has changed across generations. Look for:

Upward trend: agents are improving, the evolution cycle is working
Plateau: training tasks may need diversification, or capability limits have been reached
Regression: roll back to a previous version and investigate

Model Merging

When multiple specialized agent versions exist, model merging can combine their strengths:

# View merge candidates
uv run gaius-cli --cmd "/evolve merge candidates" --format json

# Trigger a merge
uv run gaius-cli --cmd "/evolve merge" --format json

# View lineage
uv run gaius-cli --cmd "/evolve lineage" --format json

Model lineage tracking records the ancestry of each merged version, enabling traceability from the final model back to its training data and parent versions.

Keyboard shortcuts

Gaius