Skip to content

Multi-Agent Execution

Tessera v0.3.0 supports true parallel multi-agent execution using asyncio for concurrent task processing.

Overview

Multi-agent execution allows multiple AI agents to work on different tasks simultaneously:

  • Parallel Processing: Tasks execute concurrently up to max_parallel limit
  • Dependency Management: Tasks wait for dependencies before execution
  • Intelligent Assignment: Best agent selected based on capabilities
  • Progress Tracking: Real-time monitoring of task status

How It Works

1. Task Decomposition

The supervisor agent analyzes your objective and breaks it into subtasks:

# Supervisor decomposes: "Build a FastAPI service"
subtasks = [
    "Design API endpoints and models",
    "Implement core business logic",
    "Add authentication middleware",
    "Write comprehensive tests",
    "Create API documentation",
]

2. Dependency Resolution

Tasks are organized with dependencies:

tasks = [
    Task(id="t1", description="Design API", dependencies=[]),
    Task(id="t2", description="Implement logic", dependencies=["t1"]),
    Task(id="t3", description="Add auth", dependencies=["t2"]),
    Task(id="t4", description="Write tests", dependencies=["t2"]),
]

3. Parallel Execution

Tasks execute concurrently when dependencies are met:

Iteration 1: [t1] executes
Iteration 2: [t2] executes
Iteration 3: [t3, t4] execute in parallel  # Both depend on t2

4. Progress Monitoring

Track execution in real-time:

progress = executor.get_progress()
# {
#   "queue": {"pending": 2, "in_progress": 2, "completed": 1},
#   "agent_pool": {"available": 1, "busy": 2},
# }

Configuration

Enable multi-agent mode in your config:

# config.yaml
workflow:
  max_parallel: 3  # Max concurrent agents
  max_iterations: 10  # Max execution loops

agents:
  definitions:
    - name: supervisor
      model: gpt-4
      capabilities: [orchestration]

    - name: python-expert
      model: gpt-4
      capabilities: [python, backend]

    - name: test-engineer
      model: gpt-4
      capabilities: [testing, quality]

CLI Usage

# Standard execution (uses multi-agent if multiple agents defined)
tessera main "Build a REST API with authentication"

# Monitor progress
tessera status  # Coming in v1.1

Quality Monitoring

Built-in quality monitoring prevents infinite loops:

  • Loop Detection: Identical outputs detected via hashing
  • Improvement Tracking: Coverage and quality score monitored
  • Automatic Termination: Stops if no improvement for N iterations
monitor = QualityMonitor(
    min_coverage_improvement=0.05,  # Require 5% coverage gain
    max_iterations_without_improvement=3,  # Stop if stuck
)

Best Practices

Agent Specialization

Define agents with specific capabilities:

agents:
  definitions:
    - name: backend-dev
      capabilities: [python, fastapi, databases]
      model: gpt-4

    - name: frontend-dev
      capabilities: [javascript, react, css]
      model: gpt-4

Task Dependencies

Supervisor automatically infers dependencies:

  • API implementation depends on design
  • Tests depend on implementation
  • Documentation depends on completed code

Parallelism Limits

Set max_parallel based on:

  • API rate limits: Stay within provider limits
  • Cost control: More parallel = higher concurrent cost
  • Quality: Too much parallelism can reduce coordination

Recommended: max_parallel: 3-5

Metrics & Observability

All multi-agent execution is tracked:

# Metrics stored in SQLite
~/.local/share/tessera/metrics.db

# Includes:
- Task assignments and completions
- Agent performance and success rates
- Cost per task and agent
- Duration and iterations

See Observability Guide for details.

Next Steps