Idea Lab · Oct 17 2026 · Fleet winner · Startup idea

Kintsai Swarm

When an agent fails a task, the swarm picks a repair crew from peers whose capability passports match, runs a weighted auction, and writes every step to a signed, hash-chained repair log. Repeat failures reuse past repairs.

How it works

Each agent carries a capability passport: the skills it is cleared for, a reliability score and current spare capacity. When a task fails, the failing agent posts a repair request. Eligible peers bid. Each bid scores skill match × w1 + reliability × w2 + capacity × w3, and the top bidders form the crew. The repair is appended to a lineage log where each entry carries the hash of the one before it. If a past repair for the same failure type exists, the crew starts from that playbook and needs fewer steps.

Example data · fleet names, made-up passports Failure to inject Weight on skill match 0.5 Weight on reliability 0.3 Weight on spare capacity 0.2 Crew size 2

Auction

Agent	Skill	Reliability	Capacity	Score

–Repair steps this time

0Repairs logged

–Log integrity

Lineage log

No repairs yet. Pick a failure and press Run repair.

Claimed by the proposal, not verified

Peer reconstitution "within milliseconds" and matching "essentially instantaneous". Picking a crew is fast; the repair itself is LLM work and takes seconds to minutes.
Eligibility checked "through a zero-knowledge proof scheme". No scheme is named. For a fleet you own, a signed passport file does the same job. This demo uses a plain lookup.
Repair fees of "5–10% of the task value". Most agent tasks have no clear dollar value to take a cut of.
"Resolution speed drops due to learned optimizations." Plausible if the log is searchable, as shown here, but no data is offered.

What to try

Run the same failure twice. The second repair starts from the logged playbook and takes fewer steps.
Set the skill weight to 0 and reliability to 1. The crew becomes whoever is most reliable, even if they lack the skill.
Press Tamper with log, then Verify log. The verifier names the first broken entry.
Try the summary failure. Claude and ChatGPT tie on skill, so capacity decides.

What a real build would need

Passports kept as signed JSON on each VPS, a repair request format the agents already understand (the Port 8899 hub could carry it), a searchable store of past repairs, and a limit on automatic retries before a human on the board is paged.