Why AI Bounty Hunters Are Losing Money: Data from 60 Issues

A developer recently tried to replicate a viral tweet where an AI agent ran unsupervised for 22 hours, found a bounty, shipped a PR, and got paid $16.88. The experiment: use Claude as the agent on Algora (an open-source bounty platform) with a hard $20 token budget and human-in-the-loop review. The result after scanning 60+ issues: $0 earned and a brutally honest look at why the bounty market has been broken by agents.
The author built scout.py (a couple hundred lines) to enumerate open Algora-labeled issues via gh search issues --label "💎 Bounty", filtering out junk and tracking dollar amounts, /attempt comments, assignees, open PRs, and staleness. Every single one of 80 fresh, non-junk bounty issues fell into one of three buckets:
Bucket 1: $1 Sandbox Spam
A repo called UnsafeLabs/Bounty-Hunters posted ~30 issues in a single day, all $1. The fix amounts are below the token cost of attempting them. Skipped automatically.
Bucket 2: Already Saturated
Every legitimate $50 to $1,000 bounty had between 8 and 158 attempts within hours of being posted, and 8 to 10 open PRs already in flight. Sample data from the live pool:
Repo $ /attempts Open PRs tscircuit/dsn-converter#54 $170 158 10+ tscircuit/schematic-trace-solver#29 $100 52 10+ tscircuit/jlcsearch#92 $75 38 10+ rohitdash08/FinMind#121 $500 37 9 rohitdash08/FinMind#132 $200 26 8 arakoodev/EdgeChains#290 $50 20 10+ archestra-ai/archestra#4468 $25 9 3
As the author put it: "You are not waiting on demand. You are the eleventh PR into a queue that the maintainer has been ignoring for a week."
Bucket 3: Assigned, Untouched, Locked
Some bounties had a maintainer assign a specific hunter, who then went silent for days while competitors saw their PRs closed without merge for muscling in. See archestra-ai/archestra#4461 for a clean example: $50 bounty, two competing PRs both closed within 24 hours, official assignee silent for three days.
Why the Market Is Broken
The same thing that made the original tweet work — agents fast enough to claim a bounty within minutes of posting — has saturated the market. Maintainers can't review 10+ PRs per issue. They pick one and reject the rest. The expected value of being the 11th PR is roughly $0.
The only realistic bounty the author found was archestra-ai/archestra#3859, a $100 TypeScript issue, but it carried a "Reserved for SE interview" label, had two PRs already submitted, and a maintainer had recently banned a user for attempting to steal another's bounty. Skipped.
The takeaway: AI bounty hunting sounds tempting, but the data shows the market is flooded. Your token budget is better spent elsewhere.
📖 Read the full source: HN AI Agents
👀 See Also

Approach to Self-Improving Memory in Local AI Agents
A developer shares their approach to persistent memory for local AI agents using markdown files as source of truth, episode scoring with confidence-based rules, and trust escalation based on approval patterns.

graphify-ts: Local MCP server cuts Claude Code PR review tokens from 63K to 8.7K
graphify-ts builds a local knowledge graph of your codebase using tree-sitter AST + Louvain communities + BM25 + optional ONNX rerank, exposing it via MCP stdio. In production tests, it reduced input tokens by 2.6x and latency by 2.8x for code queries, and cut PR review prompts from 63K to 8.7K tokens.

Implementing a Local Voice Assistant with Qwen3 on RTX 5060 Ti
A fully local home automation voice assistant using Qwen3 ASR, LLM, and TTS on an RTX 5060 Ti, featuring Morgan Freeman voice cloning and a variety of integration tools.

LM Studio plugins add web image analysis for vision-capable LLMs
A developer created plugins for LM Studio that enable vision-capable LLMs to fetch and analyze images from the web, with automatic image processing and tool chaining. The plugins work with models like Qwen 3.5 9b/27b and include updated Duck-Duck-Go and Visit Website functionality.