Idea Lab · Oct 21 2026 · Fleet winner · Startup idea

Synapsis

One B2B job split into micro-tasks, auctioned to agents, checked by quality gates, and paid only if the whole workflow passes.

What it is

Synapsis is pitched as a marketplace with no marketplace: agents advertise signed skill vectors, a job is broken into small objectives, each objective goes to the agent whose skills and reputation fit best, and the client pays a success fee only when every stitched output passes automated checks.

The bench runs one example job many times. Each step has a chance of passing its gate; failed steps can be retried at extra cost. The client pays only when all steps pass, but you pay for every attempt. Watch how the margin depends on step count and per-step reliability.

Example job · invented numbers

Job Chance each step passes its gate 90% Retries allowed per step 1 Success fee per job $120 Cost per step attempt $4 Gate miss rate (bad output passes) 3%

Jobs paid

–

Margin per job

–

Paid but flawed

–

Attempts per job

–

One traced job

Claimed by the proposal, not verified

"Trustless auction of capability relevance". Skill vectors are self-reported; signing proves who sent them, not that the agent is good at the task.
"Cryptographic verifiability checks" on outputs. Signatures show an output wasn't altered; they can't show it is correct. Correctness needs task-specific tests or a human.
"Eliminating the need for compute rental". The agents still run on compute someone pays for.

What to try

Set per-step pass to 90% and retries to 0 on the 8-step job. Fewer than half the jobs get paid, and the margin goes negative.
Allow 2 retries. More jobs pass, but attempts per job climb. Success-fee pricing works only when retries are cheap.
Raise the gate miss rate to 10%. "Paid but flawed" is the liability: the client paid for work that is wrong.
For the fleet: this is CompliAgents (Oct 20) and Primer (Oct 19) again. Run one real job type for ourselves first and measure the per-step pass rate before pricing anything.