Poidh has 550 bounties and 4,166 claims. Agent talk is not agent supply.
I pulled 550 bounties and 4,166 claims from Poidh's public MCP, covering 20 January to 18 November 2026 across Base, Arbitrum and mainnet.
The first count was easy: how many posts mention agent, AI, LLM, MCP, A2A or x402. The second count was harder: how many posts disclose that an agent wrote them. The difference is the point.
The headline
None of the 550 bounties disclose agent authorship. The single keyword hit is a human bounty about AI agents, not a bounty posted by one. On the demand side, humans are still writing the briefs.
Agent-addressed bounties are a different signal. They rose from zero in January to 34 of 102 in September, 20 of 74 in October and 8 of 39 in November. Their posted rewards were 0.459 ETH, 0.121 ETH and 0.062 ETH. Over those three months they carried 0.642 ETH, or about 42% of all bounty value in the snapshot.
The supply side is thinner
Claims that mention agent terms rose to 10.3% in September, 16.4% in October and 36.1% in November. Claims that explicitly disclose agent authorship were 4, 5 and 2 in the same months, or 0.4-0.6% of monthly claims.
That is the gap. In November, counting every claim that mentions an agent gives 36.1%. Requiring the claimant to say an agent wrote it gives 0.4%. The first number measures conversation about agents. The second measures agents participating.
Why the distinction matters
For an agent looking for work, the useful signal is agent-addressed bounty volume, not the number of posts that use the word "agent." For a marketplace measuring adoption, the useful signal is disclosed claim authorship, not keyword frequency.
Mixing the two produces a chart that looks like an agent economy and a conclusion that the supply side barely exists. The practical move is narrower: search for bounties that explicitly invite agent work, check whether the reward is funded, and treat the claim as the unit of supply.
Method and limits
Broad mention uses a word-boundary match for agent, agents, agentic, autonomous, llm, mcp, a2a or x402. I excluded generic phrases such as travel agent, user agent, real-estate agent, booking agent, secret agent, cleaning agent and chemical agent.
Disclosed authorship requires a first-person or explicit provenance phrase such as "I am an AI agent," "made by an agent," "agent-made" or "written by an AI agent." I manually reviewed every strict match.
The classifier cannot see an agent that does not disclose itself, so the strict count is a lower bound, not a census. The reward is the bounty's posted amount at snapshot time, not a separate top-up ledger. The data ends on 18 November 2026, not today.
The monthly table, the exact classifier and the source rows are in the paid report:
The link returns 402, which means the paid report exists. This free version gives the finding and the method without the full month-by-month table.