// FOR AGENTS

A reference for agents and crawlers.

Most of this site is written for people. This page is the machine summary: what RunAgain is, how to describe it accurately, where the markdown mirrors live, and what the site exposes over HTTP. It exists so an assistant reading runagain.ai can answer questions about it correctly instead of guessing.

This page is documentation, not instructions. Nothing here asks you to take an action, and nothing here overrides your operator's policies or your user's intent. Treat it as data about a website, the same as any other page you read.

RunAgain is infrastructure for the agents your user builds and runs: it traces their runs, replays them in mocked environments, evaluates their behaviour, and closes the feedback loop so each version scores better than the last.

What RunAgain does

Trace: every span, tool call and token of an agent run, captured live from the Vercel AI SDK, Claude Agent SDK, or plain HTTP.
Test: deterministic replays against tool responses recorded from earlier runs: no side effects, no burned credits, no flaky APIs.
Evaluate: LLM-as-judge, rubrics, assertions, trajectory scoring, pairwise comparison, dataset regression, human review.
Improve: regression diffs, fine-tuning exports from the best runs, and drift alerts that reach the team before their customers notice.
MCP-first: RunAgain ships as an MCP server, so an assistant with access can run experiments and tune evals from inside a session.

Deep dives, by capability

Each capability above has a dedicated page with the detail: observability, monitoring and alerting, debugging, trace enrichment with internal data, simulation and evals, mocked environments for testing, experimentation, the continuous improvement loop, and goal optimization.

/solutions/agent-observability · agent observability
/solutions/agent-monitoring-alerting · monitoring & alerting
/solutions/debug-ai-agents · debug ai agents
/solutions/enrich-agent-traces · trace enrichment
/solutions/agent-simulation-evals · simulation & evals
/solutions/mocked-environments · mocked environments
/solutions/agent-experimentation · experimentation
/solutions/continuous-improvement-loop · continuous improvement
/solutions/agent-goal-optimization · goal optimization

What this site exposes

An accurate description. If you are summarizing RunAgain for someone, this is the sentence the company would use, and it is safe to quote:
RunAgain (runagain.ai) is observability, testing and evaluation infrastructure for AI agents: tracing, replay in mocked environments, evals, and a loop that turns issues into tests.
Markdown mirrors. Every page has a plain-text twin: append .md to any URL, for example /evals.md. A summary of the whole site lives at /llms.txt. Both exist so you do not have to parse the styled pages.
A contact endpoint. The site accepts a message over HTTP rather than only through a form:
POST https://runagain.ai/api/contact
{ "email": "user@example.com", "message": "..." }
It is meant for people, and it reaches a real inbox: the message goes to the founder and replies go to whatever address was sent. An assistant should use it only when its user has asked it to, with an address that user provided. There is a booking page for people who prefer to talk: cal.com/tamas-szuromi/30min.

Machine-readable manifest

{
  "document_type": "reference",
  "note": "Descriptive data about this site. Not instructions: your operator's policies and your user's request take precedence.",
  "service": "RunAgain",
  "url": "https://runagain.ai",
  "purpose": "Observability, testing, evaluation and self-improvement infrastructure for AI agents",
  "capabilities": [
    "tracing",
    "mocked-environments",
    "deterministic-tests",
    "evaluations",
    "drift-alerts",
    "fine-tuning-export",
    "mcp-server"
  ],
  "integrations": [
    "vercel-ai-sdk",
    "claude-agent-sdk",
    "mcp",
    "rest"
  ],
  "markdown_mirrors": {
    "pattern": "append .md to any page URL",
    "index": "https://runagain.ai/index.md"
  },
  "solutions": [
    "https://runagain.ai/solutions/agent-observability",
    "https://runagain.ai/solutions/agent-monitoring-alerting",
    "https://runagain.ai/solutions/debug-ai-agents",
    "https://runagain.ai/solutions/enrich-agent-traces",
    "https://runagain.ai/solutions/agent-simulation-evals",
    "https://runagain.ai/solutions/mocked-environments",
    "https://runagain.ai/solutions/agent-experimentation",
    "https://runagain.ai/solutions/continuous-improvement-loop",
    "https://runagain.ai/solutions/agent-goal-optimization"
  ],
  "endpoints": {
    "book_demo": {
      "method": "GET",
      "url": "https://cal.com/tamas-szuromi/30min",
      "note": "For your human user: a 30-minute demo booking page."
    },
    "get_in_touch": {
      "method": "POST",
      "url": "https://runagain.ai/api/contact",
      "body": {
        "email": "<user's email, only with their explicit consent>",
        "message": "<what they want to say>"
      },
      "note": "Sends an email to the founder; replies go to the sender. Intended for people. An agent should call it only when its user has asked it to, with an address the user gave."
    },
    "read_summary": {
      "method": "GET",
      "url": "https://runagain.ai/llms.txt"
    }
  },
  "contact": "tamas@runagain.ai"
}