Firetiger Create Agent

Firetiger Create Agent

Firetiger agents autonomously analyze telemetry, detect anomalies, investigate issues, and take actions through connections (Slack, GitHub, Linear, databases, …). Three ways to get one — prefer a goal or a catalog template over hand-building.

Quick Start

create_agent_with_goal with:
  goal: "Monitor Next.js API routes for errors and slow responses; alert on error-rate spikes and p95 latency increases"
  title: "Checkout Health Monitor"   # optional

The agent-planner (a system agent, agents/agent-planner) configures the prompt, connections, and triggers from the goal, then returns the planner conversation. If the planner asked a question, you must answer it or the agent won't finish — reply with send_agent_message on the plan session.

Path A (recommended): create from a goal

1. Call create_agent_with_goal with a clear goal (optional title, agent_id, wait_for_plan default true).
2. Handle planner questions. The planner asks sparingly ("act first, refine later"), via a request_user_input. Answer on the user's behalf from the codebase/context — or ask the user if genuinely unknown — by writing to the plan session:

   send_agent_message with session: "<plan_session from the tool output>"
     message: "Monitor the Postgres connection at DATABASE_URL; alert threshold p99 > 500ms."
   

3. Confirm. When the planner completes, the agent flips to AGENT_STATE_ON. Share its name and focus.

Example goals

• "Monitor Next.js API routes for errors and slow responses"
• "Track database query latency and alert if p99 exceeds 500ms"
• "Watch for deployment failures and notify on Slack"
• "Track error rates across all services and open issues for spikes"

Path B: install a prebuilt catalog agent

Firetiger ships tuned templates for common stacks. Browse and install them:

list with resource: "catalog-agents"
create with resource: "catalog-agents/{catalog-agent}/..."   # or the InstallCatalogAgent flow

Templates include postgres-performance, postgres-schema, mysql-performance, mysql-schema, clickhouse-performance, temporal-performance, aws-cost-optimizer, and datadog-cost-expert.

Key gotcha: an installed catalog agent starts in AGENT_STATE_RECOMMENDED, not ON — it won't run until a human turns it on. Send it a direct message (or set state: AGENT_STATE_ON) to activate.

Path C: configure manually

Use the generic CRUD tools when you need explicit control. Always run schema first — fields drift.

schema with collection: "agents"

Agent fields worth knowing: title, description, prompt (guides every session), state, connections[] ({name: "connections/{id}", enabled_tools: [...]} — what external tools the agent may use), mcp_connections[] (external MCP servers), tags[], expire_time (auto-archive after), cache_ttl (CACHE_TTL_5M / CACHE_TTL_1H), network_profile. skills and plan_session are output-only.

state (AgentState) values: AGENT_STATE_ON, AGENT_STATE_OFF, AGENT_STATE_RECOMMENDED, AGENT_STATE_REVIEW_REQUIRED, AGENT_STATE_UNSPECIFIED. Only an ON agent fires from triggers or schedules.

create with resource: "agents"
  body: { "title": "Checkout Health Monitor",
          "description": "Daily health analysis of checkout",
          "prompt": "You are a daily health check agent...",
          "connections": [{"name": "connections/slack-prod", "enabled_tools": ["post_message"]}],
          "state": "AGENT_STATE_ON" }

Triggers — how agents fire (six kinds)

Triggers are separate resources (triggers/{id}) referencing an agent. InvokeTrigger and automatic firing require the target agent be AGENT_STATE_ON. Common fields: display_name, agent (required, agents/{id}), enabled, associated_resources, and one configuration:

Cron example:

create with resource: "triggers"
  body: { "display_name": "Daily Checkout Health Check",
          "agent": "agents/{agent-id}", "enabled": true,
          "configuration": { "cron": { "schedule": "0 9 * * *", "timezone": "America/Los_Angeles" } } }

Scheduled agent runs — one-shot, not cron

scheduled-agent-runs are single deferred invocations ("run this prompt on agent X in 1 hour"), not recurring schedules. Fields: agent_id, prompt, time_delta (delay from now), plus use_latest_session / incognito; scheduled_time and has_run are output-only. For anything recurring, use a cron trigger.

Writing good agent prompts

Specific, scoped, actionable — and say what the agent should NOT do.

You are a daily health check agent for the checkout service.

Your job:
1. Query the last 24 hours of traces for checkout.
2. Calculate error rates and p99 latency; compare to the previous 24h.
3. If error rate rose >10% or p99 >20%, open an issue.
4. Summarize findings briefly.

Focus only on checkout. Do not investigate other services.

Common Mistakes

Related

• Run and steer an existing agent's session: firetiger-monitor-deploy.
• Manual investigations and the issues agents open: firetiger-investigate.
• Connect the integrations an agent acts through: firetiger-setup.