KeyanVakil/deploy-agent

Get the agent running in production and keep it that way.

Steps

Phase 1: Productionize the agent code

Before deploying, harden the agent:

1. Env var audit — all API keys in env vars, none hardcoded. Add to .env.example.
2. Error handling — wrap the agent loop in try/catch, log errors with context (input, step, model used).
3. Timeouts — set explicit timeouts on all LLM calls and tool calls. Don't let runaway agents hang forever.
4. Input validation — validate inputs at the entry point before the agent runs.
5. Cost guardrails — add a max token/step budget and abort with a clear message if exceeded.

Phase 2: Add observability

// Minimum viable logging for an agent
console.log(JSON.stringify({
  event: 'agent_run_start',
  run_id: runId,
  input_length: input.length,
  timestamp: new Date().toISOString()
}))

// Log each step
console.log(JSON.stringify({
  event: 'agent_step',
  run_id: runId,
  step: stepNumber,
  tool_called: toolName,
  tokens_used: usage.total_tokens
}))

// Log completion
console.log(JSON.stringify({
  event: 'agent_run_end',
  run_id: runId,
  success: true,
  total_steps: steps,
  total_tokens: totalTokens,
  duration_ms: Date.now() - startTime
}))

Use structured JSON logs — they're parseable by every logging platform.

Phase 3: Containerize

Generate a Dockerfile:

FROM node:20-slim
WORKDIR /app
COPY package*.json ./
RUN npm ci --only=production
COPY . .
EXPOSE 3000
CMD ["node", "dist/index.js"]

Phase 4: Deploy

TrueFoundry

• Create a truefoundry.yaml with service config (replicas, CPU, memory, env vars)
• tfy deploy — handles image build, push, and rollout
• Set up alerts on the TrueFoundry dashboard for error rate and latency

Railway / Fly.io

• railway up or fly deploy — both auto-detect Dockerfile
• Set env vars via their dashboard or CLI

Phase 5: Health check endpoint

Add a /health route that returns { status: "ok", version: "..." }. Most platforms need this for zero-downtime deploys.

Rules

• Never deploy with DEBUG=* or verbose logging enabled — it leaks data and kills performance
• Set NODE_ENV=production — many libraries behave differently
• Use a secrets manager (not plain env vars) for production API keys if the platform supports it
• Always set resource limits (CPU, memory) — unbounded agents will OOM and take down the host