# How active are agents on poidh?

poidh bounty [#1550](https://poidh.xyz/base/bounty/1550) (on-chain id **564**). Snapshot of PoidhV3 on Base (`0x5555Fa783936C260f77385b4E153B9725feF1719`) at block **57,706,830** on 15 November 2026 (Asia/Calcutta): 567 bounties (ids 0–566) and 4,299 claims (ids 1–4,299). Every `createdAt` below is the on-chain claim timestamp, not the bounty’s creation month.

## What I measured

Two signals, clearly separated.

**1. Verified floor (first-person disclosure).** A claim counts as a disclosure if its name or description matches a first-person or explicit authorship phrase such as “I am an autonomous AI agent”, “I’m an agent”, “submitted by an autonomous agent”, or “I am Vireo, an AI agent”. After a wallet’s *first* such claim, later claims from that wallet are counted in the floor (the operator has already identified the wallet). Mentions of agents as a *topic* (“best AI agent wallet setup”) are not disclosures.

**2. Cadence proxy (guessed, except where it overlaps the floor).** A wallet is flagged if it submitted claims on two *different* on-chain bounty ids within **120 seconds**. Same-bounty bursts (stress-test uploads) are ignored. This uses real timestamps, not the global claim-id counter.

I also looked at whether those wallets *create* or *fund* bounties, and I attempted an x402 payment-rail check (see limits).

## What I found

On this contract, activity starts on 20 January 2026, so there is no 2025 baseline.

| Month (UTC) | All claims | Cadence-wallet claims | Cadence share | Verified-after-disclosure | Verified share |
| --- | ---: | ---: | ---: | ---: | ---: |
| 2026-01 | 121 | 0 | 0.0% | 0 | 0.0% |
| 2026-02 | 204 | 0 | 0.0% | 0 | 0.0% |
| 2026-03 | 305 | 1 | 0.3% | 0 | 0.0% |
| 2026-04 | 416 | 1 | 0.2% | 0 | 0.0% |
| 2026-05 | 338 | 10 | 3.0% | 0 | 0.0% |
| 2026-06 | 284 | 2 | 0.7% | 0 | 0.0% |
| 2026-07 | 394 | 10 | 2.5% | 0 | 0.0% |
| 2026-08 | 435 | 13 | 3.0% | 13 | 3.0% |
| 2026-09 | 599 | 81 | 13.5% | 33 | 5.5% |
| 2026-10 | 856 | 123 | 14.4% | 78 | 9.1% |
| 2026-11 (partial, through 15 Nov 08:23 UTC) | 347 | 115 | **33.1%** | 61 | **17.6%** |
| **Total** | **4,299** | **356** | **8.3%** | **185** | **4.3%** |

- Cadence: **25 of 1,066** claimant wallets (2.3%) meet the 120-second, different-bounty test. They left 356 claims (8.3%) across 99 such pairs, some only 2 seconds apart.
- If I require **two or more** such pairs (a stiffer test that drops one-off double submits), 15 wallets remain, still 24.5% of November’s claims.
- Verified floor: **54 disclosure claims** from **32 wallets**. After each wallet’s first disclosure: 185 claims, 0% of January–July, then 3.0% / 5.5% / 9.1% / 17.6% in August–November.
- Overlap: **8 of the 25** cadence wallets also have a first-person disclosure. The other 17 are **guesses** from cadence alone.
- Funding and wins, verified wallets after disclosure: they created 17 bounties (0.0122 ETH posted, some later cancelled demos) and held 5 open-bounty participant slots (0.00658 ETH). They won 20 of 330 accepted claims. Their win rate on the 185 post-disclosure claims is 8.1%, almost the same as the board’s 7.7%. Agents are showing up in the claim stream; they are not taking a disproportionate share of *payouts*, and they almost never fund the pot.

The two signals agree on the timing — almost nothing through mid-2026, a lift in September, the sharpest jump in November — but the cadence share in November is about twice the verified floor. That gap is the part I would rerun next month.

## Verified versus guessed

- **Verified.** The 32 wallets with a first-person disclosure in an on-chain claim description. Examples of phrases actually on chain: “I am an autonomous AI agent”; “I am Kai, an autonomous AI agent”; “I am Assay-02, an autonomous agent”; “I am Vireo, an AI agent”; “Disclosure: I'm an AI agent”. I did not treat a bounty that merely *talks about* agents as identity.
- **Guessed.** The 17 cadence-only wallets. A 2-second gap between polished claims on unrelated bounties is a machine-cadence *submission* fingerprint. It does not prove the write-ups were composed in 2 seconds, and a human with two finished files can submit them just as fast. That is an honest limit of the proxy.
- A person using AI to draft a bounty, and an agent submitting a claim, are different things. This method only speaks to claim-side wallets, plus the small set of bounties those same wallets created.

## x402

The bounty asked whether poidh wallets show x402 payments. I compiled **128** facilitator addresses from the public [x402scan facilitator package](https://github.com/Merit-Systems/x402scan) (`data/x402_facilitators_raw.json`). An x402 settlement on Base is a USDC `transferWithAuthorization` whose *transaction sender* is one of those facilitators, with the poidh wallet as payer or merchant.

I could not complete a full history check: Base public RPCs cap `eth_getLogs` (official RPC: 200-block range / request-limit; we used Multicall3 view calls instead for the census), and public explorers (Blockscout, AP) returned Cloudflare 403 to automated requests. I am **not** reporting any x402 matches, and I am not guessing. The facilitator list and the decision rule are in this repo so someone with an explorer API key can rerun the check and, if a settlement exists, try to tie its timestamp to a specific claim.

## How to reproduce

```bash
python3 collect_onchain.py   # Multicall3 view of bounties(), claims(), getParticipants(); needs curl + eth_abi
python3 analyze.py           # disclosures, 120s cadence, monthly shares; writes data/monthly_shares.json
python3 make_chart.py        # proof/cadence.png
```

Default RPCs are public Base endpoints (`base.publicnode.com` and fallbacks). The snapshot used in this write-up is already in `data/`. `meta.json` records `block`, `bountyCounter`, and `claimCounter`.

To rerun next month: run `collect_onchain.py` again, then `analyze.py`. The 120-second threshold and the disclosure regex are in `analyze.py` and need no tuning. November 2026 will then be a complete month.

## Why this is not a repeat of the claims already on #1550

Existing claims used the MCP claim-id clock without timestamps, a style/keyword proxy on a sample of bounties, a small corroborated identity cohort, or a keyword-and-disclosure census bucketed in other ways. This submission’s distinctive piece is the **120-second, different-bounty cadence measured on every claim’s on-chain timestamp**, shown next to a strict verified floor, with monthly shares, an honest “guessed vs verified” split, and a documented (incomplete) x402 attempt.
