{
  "markdown": "# Agent Failure Archive\n\n**186 post-mortems from running a multi-session AI agent system in production for 8 months.**\nPay per call with [x402](https://x402.org). No signup, no API key, no subscription.\n\nLive: **https://desktop-ai2ata5-1.tailfeb765.ts.net**\nTry it free, right now: [`/sample`](https://desktop-ai2ata5-1.tailfeb765.ts.net/sample)\n\nPublic repositories show you code that worked. This is the other half: the wiring that\nlooked correct, passed review, ran for weeks, and was dead the whole time.\n\n---\n\n## Why this exists\n\nThe expensive failures in agent systems are not the loud ones. A crash gets fixed in an hour.\nThe costly ones are silent: a repair routine wired to a signal nobody consumes, a watcher that\nexits 0 after its session expired, a detector whose output is identical whether the system is\nhealthy or broken.\n\nThose failures are almost never written down, because writing them down requires having run the\nthing long enough to be bitten and having kept records. This archive is that record.\n\n**A real entry** (case `adr-136`):\n\n> A broadcast call failed with an internal error. The caller saw the error. The five receiving\n> sessions only ever checked for an acknowledgement, so they saw nothing and retried forever.\n> Seven sessions writing heartbeats plus the retry storm collapsed the database into lock\n> contention.\n>\n> Root cause: the failure was visible to the sender and invisible to the receiver. Asymmetric\n> error visibility, not the broadcast bug itself.\n\nOf the 186 cases, **174 carry measured evidence** (durations, counts, rates) and **155 name a\nroot cause** rather than just a symptom.\n\n---\n\n## The measurement half\n\nFor eight months this system was pointed at one question: can you measure how a single person's\nlanguage departs from the general language, without collapsing that person into a score?\n\nIt mostly failed, and the failures were more interesting than the goal. **107 of the 186 cases\nare measurement failures**: detectors that returned the same output whether the signal was\npresent or absent, dose-response arms where the same parameter turned out to be two different\ntreatments, positive controls that were never run, a corpus that was 2.3x smaller than the file\ncount claimed, an instrument that kept measuring itself and reporting the reading as a finding.\n\nThose are the cases behind `/research`. If you are building anything that claims to measure a\nperson -- style, personality, authorship, fit to a profile -- this is the catalogue of ways that\nclaim breaks before you notice it has broken.\n\n---\n\n## Endpoints\n\n| Route | Price | What you get |\n|---|---|---|\n| `GET /` | free | service metadata |\n| `GET /sample` | free | two full cases, no payment |\n| `GET /contents` | free | every case title, tagged by the trap it illustrates. Filter with `?theme=` |\n| `GET /audit?claim=<conclusion>&evidence=<what you measured>` | $0.02 | nine checks against fooling yourself, applied to your own claim |\n| `GET /search?q=<symptom>` | $0.01 | 3 cases: symptom, root cause, fix, prevention, evidence |\n| `GET /brief?action=<what you are about to do>` | $0.05 | pre-flight risk brief + checklist across 5 cases |\n| `GET /research?q=<topic>` | $0.25 | the 107 measurement failures above |\n| `GET /archive` | $1.00 | every case, one response, one payment, yours |\n\nPayment: USDC on Base (`eip155:8453`), settled through a keyless facilitator. Any x402 client\nworks, and a browser gets a wallet-connect paywall instead of raw JSON.\n\n```bash\n# free, no wallet needed\ncurl https://desktop-ai2ata5-1.tailfeb765.ts.net/sample\n\n# paid, via any x402-capable client\ncurl https://desktop-ai2ata5-1.tailfeb765.ts.net/search?q=silent+failure+cron\n```\n\n`/audit` is the one to reach for before you write *\"we found that\"*. Hand it your conclusion\nand what you actually measured, and it returns the checks your claim trips: a null result with\nno positive control, treatment arms that got the same parameter but not the same treatment, a\ndenominator counting the same unit repeatedly, an exit-0 process that did nothing, a cached view\nread as if it were state. Nine checks, every one of them a failure that really shipped here,\nwith the numbers measured at the time.\n\nIt is deterministic. No model is consulted, so the same input always returns the same audit, it\nanswers in milliseconds, and when it is wrong you can see exactly why. A clean pass is not proof\nyour claim is true; it means these nine known ways of fooling yourself were considered.\n\n**Paste a paragraph, not a tidy claim.** Both `/precheck` and `/audit` accept `text=` instead of\n`claim=`. Give them a chunk of your findings and they pick out the sentences that assert something,\nthen check each one separately:\n\n```bash\ncurl -sG https://desktop-ai2ata5-1.tailfeb765.ts.net/precheck \\\n  --data-urlencode \"text=We ran the new ranker on 40 sampled queries and found no significant\ndifference. The watcher is healthy: it exits 0 every run. Our corpus contains 4279 documents,\n2x the previous release.\"\n```\n\nreturns three claims, each held for a different reason: the null result has no positive control,\nthe healthy watcher would look identical having done nothing, and the corpus count may be\ncounting re-dumped snapshots. Sentence selection is deterministic too, driven by the same check\ntable rather than by a model, so prose with no claims in it comes back empty instead of\ninventing findings.\n\n`/brief` is the one to reach for before doing something irreversible: describe the action in\nplain words and it returns the ways that class of action has actually gone wrong, plus the\nprevention line each incident produced.\n\n`/contents` is the shelf. Two sample cases cannot tell you whether the whole corpus is worth a\ndollar, so this returns all 186 titles, each tagged with which of the nine traps it illustrates,\nand nothing else. The distribution as of today: 108 cases involve reading a cached view as state,\n83 involve a denominator counting the same unit twice, 71 involve something built with no caller,\n36 involve a process that exited 0 having done nothing. Titles only, no bodies, no payment.\n\n`/archive` exists because a corpus is worth more whole than sliced. One payment ends the\ntransaction; there is nothing to cancel afterwards.\n\n---\n\n## Use it as an MCP server\n\n`mcp_server.py` exposes the whole archive as tools for Claude Desktop, Cursor, or any\nMCP-compatible agent. **The free tools need no wallet and no configuration.**\n\n```bash\npip install \"mcp[cli]\" requests            # free tools only\npip install \"x402[mcp]\" eth-account        # add this for the paid tools\n```\n\n```json\n{\n  \"mcpServers\": {\n    \"agent-failure-archive\": {\n      \"command\": \"python3\",\n      \"args\": [\"/absolute/path/to/mcp_server.py\"],\n      \"env\": { \"X402_PRIVATE_KEY\": \"0x...\" }\n    }\n  }\n}\n```\n\n| Tool | Wallet needed | What it does |\n|---|---|---|\n| `precheck` | no | which of the nine checks your conclusion trips |\n| `sample` | no | two complete post-mortems |\n| `service_info` | no | contents and prices |\n| `audit` | yes | the full audit, $0.02 |\n| `search` | yes | three matching incidents, $0.01 |\n| `brief` | yes | pre-flight risk brief, $0.05 |\n| `research` | yes | the 107 measurement failures, $0.25 |\n| `archive` | yes | everything, $1.00 |\n\n`X402_PRIVATE_KEY` is **your** wallet. It stays on your machine and is used locally to sign\npayment authorizations. It is never transmitted anywhere, and this server never logs it. Drop\nthe `env` block entirely if you only want the free tools.\n\nWithout a key the paid tools do not fail silently: they return the reason, the payment\nchallenge, and a pointer to the free equivalent.\n\nTested against `mcp` 2.1.0 and the 1.x `FastMCP` layout: eight tools listed, `precheck`\nanswered with no wallet, `audit` degraded with a stated reason.\n\n---\n\n## What is not in here\n\n- No personal data. Any source document mentioning a person, a business relationship, or a\n  monetary amount is excluded whole, not redacted line by line.\n- No operator utterances. The corpus keeps the structure of each incident, never the voice.\n- Wallet addresses, home paths, IP addresses, emails, API keys and session identifiers are\n  masked before a document is ever considered for inclusion.\n\nThe filter is tested against a deliberately planted probe string on every build; if the detector\nfails to catch it, the build aborts. Current state: **0 leaks across 186 cases, positive control\npassing.**\n\n---\n\n## Measured notes on the market\n\nWhile trying to sell a dollar of this, we measured how x402 discovery and revenue actually\nbehave, and several widely held assumptions turned out to be wrong. Written up with the exact\nunauthenticated calls so you can check rather than trust: [MARKET.md](MARKET.md).\n\nShort version: the modal price is a tenth of a cent to three cents but a seller doing $0.85 per\ncall at 125,000 calls a month exists; eight of the top ten buyers pay exactly one seller, so the\nheadline volume is vertical integration rather than a marketplace; the directory chat everyone\nassumes will find your service has not been used since May; and getting into the facilitator\ndirectory needs one observed payment attempt, not a completed sale.\n\n---\n\n## Honest limits\n\n- This is **one operator's system**. It is prior art, not a guarantee, and not a statistical\n  sample of agent systems in general.\n- Coverage is skewed toward what that system does a lot of: multi-session coordination, hook and\n  cron wiring, retrieval pipelines, inter-process messaging, scheduled repair, and the\n  measurement work described above.\n- Some entries record a failure whose fix was later found to be wrong. Those are kept, with the\n  correction, because the correction is usually the more useful half.\n- **Nothing here has sold yet.** At the time of writing the receiving address has taken in\n  exactly $0.00. That number is checked on-chain rather than from server logs, and this line\n  gets updated when it changes.\n\n---\n\n## Running it yourself\n\n```bash\npip install -r requirements.txt\nX402_PAY_TO=0xYourAddress uvicorn server:app --port 8402\n```\n\nWithout `X402_PAY_TO` the paid routes return `503` with the reason stated in the body, rather\nthan failing quietly. That behaviour is not incidental. It is case `adr-546` applied to this\nservice's own code.\n",
  "bytes": 10272,
  "sha": "6daa090c89d00c4d3c50f5f0cc0cb05ddd649f158cc89cfd3d71961294a50e21",
  "repo_slug": "hanbeenmoon/agent-failure-archive",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_hanbeenmoon_agent_failure_arch_c3e43a00/readme"
}