Dark Factory — Autonomous AI Agents

Create agent and ai_task sub-components to run autonomous AI agents inside FastMCP. Tasks execute on a schedule using Claude AI with access to all MCP tools.

How It Works

1. Agent
Defines AI model, API key, max iterations

2. Task
Defines what to do, schedule, mode, notifications

3. Link
ai_task-agent link connects task to its agent

4. Run
Background loop executes tasks, sends notifications

Agent Sub-Component

Each agent is an AI model configuration. Multiple tasks can share one agent. The agent defines who does the work.

Attribute	Default	Description
`ai_api_key` REQUIRED	—	API key for the AI provider (Anthropic or OpenAI-compatible)
`ai_model`	`claude-haiku-4-5-20251001`	Claude model ID. Haiku = cheap + fast. Sonnet = smarter for complex tasks.
`agent_ai_provider`	`anthropic`	`anthropic` or `openai-compatible` (for vLLM, Ollama, etc.)
`ai_base_url`	(empty)	Custom API base URL for OpenAI-compatible providers. Leave empty for Anthropic.
`ai_max_iterations`	`10`	Max tool-use loop iterations per task run. Higher = more thorough but costs more tokens.
`agent_enabled`	`true`	Master switch. Set to `false` to stop all tasks linked to this agent. Survives pod restarts (persisted in ENV).

ai_task Sub-Component

Each task defines what the agent should do. Link to an agent via ai_task-agent. Chain tasks via ai_task-ai_task.

Attribute	Default	Description
`task_description` REQUIRED	—	Natural language instruction. Write it like you'd tell a human operator.
`cron_expression`	(empty)	Standard 5-field cron. Empty = run once on pod start. `/30 * * ` = every 30 min `0 9 * 1-5` = weekdays 9am `0 0 * * 0` = Sunday midnight
`agent_mode`	`monitor`	monitor = read-only tools only (safe for production) act = read + write tools (INSERT, DELETE, publish, scale, sync)
`retry_count`	`0`	How many retries on failure. `0` = no retry, fail and stop.
`retry_delay_seconds`	`60`	Seconds to wait between retries.
`notify_webhook`	(empty)	Webhook URL for JSON notifications. Works with Slack, Teams, PagerDuty, or any generic webhook.
`notify_email`	(empty)	Email address for HTML reports with status badge, metadata table, and full output.

Cron Schedule Examples

cron_expression	When it runs
`/5 * * *`	Every 5 minutes (urgent monitoring)
`/30 * * *`	Every 30 minutes (standard monitoring)
`0 9 * * 1-5`	Weekdays at 9am UTC (daily report)
`0 0 * * 0`	Sunday midnight (weekly review)
`0 0 1 * *`	First day of month (monthly compliance)
`(empty)`	Run once on pod start, then never again

Sub-Component Links

Link Type	From → To	Purpose
`ai_task-agent`	ai_task → agent	Which AI runs the task. Each task can target a different agent (cheap Haiku for data gathering, smart Sonnet for analysis).
`ai_task-ai_task`	ai_task → ai_task	Task chaining. On success, triggers next task with previous output as context. Only first task needs cron.

Agent fallback: If a task has no ai_task-agent link, it automatically uses the first available agent. In single-agent setups, you don't need to link every task — the fallback handles it.

Task Chaining

Link tasks via ai_task-ai_task to create pipelines. On success, the next task runs automatically with the previous result as context.

health-check

cron: */15 * * * *

agent: haiku

→

diagnose-report

chain (no cron)

agent: sonnet

→

remediate

chain, mode: act

agent: sonnet

Only the first task needs a cron. Each task in the chain can use a different agent (cost optimization).

Chain rules:

Only the first task needs cron_expression — chained tasks trigger automatically on success
Each task can use a different agent (cheap for data, smart for analysis)
Chain stops if a task fails (unless retry_count > 0)
Chained tasks receive the previous task's full output as context

Examples

Prometheus + Loki Monitoring

"Query Prometheus for pods with CPU > 80% or memory > 90%. Check Loki for error logs. Summarize findings."

mode: monitor cron: */30 * * * *

Weekly Stack Validation

"Run all stack-agent validation tests. Report failures with test name, component, and error details."

mode: monitor cron: 0 9 * * 1 retry: 2 (60s delay)

Database Maintenance (act mode)

"Check PostgreSQL for tables with > 1M rows and no recent VACUUM. Run ANALYZE on tables that need it."

mode: act cron: 0 3 * * 0

One-time Security Audit

"Check OPA compliance for all namespaces. Report security violations grouped by severity."

mode: monitor cron: (empty) = run once

Enable / Disable

Method	How	Persists?	When to Use
Permanent	Set `agent_enabled=false` on agent sub-component in Stacktic UI. Regenerate + `kubectl apply -k`.	Yes	Long-term disable, maintenance windows
Runtime	Call `dark_factory_toggle_agent("agent-name", false)` from Claude Desktop or any MCP client.	No (resets on restart)	Quick pause, debugging, cost control

Dark Factory MCP Tools

These 6 tools are registered when agent/task sub-components exist. Use them from Claude Desktop or any MCP client to monitor and control the autonomous loop.

Tool	Description
`dark_factory_status`	Status of all agents and tasks — cron, last run, cost, retry state
`dark_factory_history(task_name, limit)`	Execution history for a task (default: last 10 runs with timestamps, status, summaries)
`dark_factory_toggle_agent(name, enabled)`	Enable/disable an agent at runtime. Takes effect instantly. Resets on pod restart.
`dark_factory_run_now(task_name, follow_chain)`	Trigger immediate execution. Optionally follow the chain on success.
`dark_factory_cost`	Per-task token usage breakdown and estimated USD cost since pod start.