Multi-Agent Workflows

Kiwi Code can spawn multiple AI agents that work in parallel. Instead of one agent doing everything sequentially, you can fan out work across several agents — each handling a piece of the task concurrently.

When to Use Multi-Agent

Multi-agent workflows shine when your task can be split into independent pieces:

• Large refactors — rename a function across 50 files, with each agent handling a batch
• Batch processing — analyze, transform, or generate content for many items at once
• Parallel analysis — run code review, test generation, and documentation updates simultaneously
• Data processing — split a CSV or dataset across agents for faster throughput

How to Trigger Multi-Agent from the Web Dashboard

Multi-agent workflows are managed through the web dashboard at kiwicode.ai. You can trigger them by:

1. Using Action Graphs — pre-configured DAG workflows where each node runs a different agent
2. Batch runs — send multiple inputs to the same action, each processed by a separate agent
3. Prompting the AI — ask the AI to split work and run tasks in parallel

Example: Ask the AI to parallelize

In the web dashboard chat, you can prompt:

Split the 20 API endpoint files in src/routes/ across agents.
Each agent should add input validation and error handling.

The AI orchestrates the fan-out automatically.

Example: Batch processing

Process each row in data.csv — for each company name,
research their API and generate an integration guide.

The platform splits the CSV and assigns each batch to a separate agent.

How It Works

  You (Web Dashboard)
    │
    ├──→ Agent 1: "Add validation to routes/users.py"
    ├──→ Agent 2: "Add validation to routes/orders.py"
    ├──→ Agent 3: "Add validation to routes/payments.py"
    │
    └──→ Results merged when all agents complete

Each agent:

• Runs independently with its own context
• Can connect to its own CLI runtime (if local execution is needed)
• Reports results back to the orchestrator
• Runs concurrently with other agents

Monitoring Multi-Agent Runs

From the TUI

Use the query commands to check on runs:

/runs list --status processing
/runs get <run_id>

From the CLI

kiwicli runs list --status processing
kiwicli runs get <run_id>

From the Web Dashboard

The dashboard shows all active and completed runs with status, duration, and results.

Local Runtime in Multi-Agent

When multi-agent workflows need to execute commands on your machine, each run gets its own isolated runtime process:

~/.kiwi/runtimes/by-run/
├── run_abc123/     ← Agent 1's runtime
├── run_def456/     ← Agent 2's runtime
└── run_ghi789/     ← Agent 3's runtime

Each runtime is scoped to its own run — agents can't interfere with each other's execution. When the TUI exits, you'll be prompted to clean up any running runtimes.

Limits