EXP-001activecoordination
Swarm routing
How should messages route through a swarm as it grows? Compares staged pipelines (the current swarm model) against gossip and quorum topologies on task completion and token cost.
the lab
Research
Agent coordination and onchain intelligence are open problems, and Yinsi treats them that way: experiments run in public, benchmark methodology is published before numbers are, and the protocol grows through written proposals. No result on this page will ever appear before its run does.
01Experiments
What the lab is running right now.
EXP-002activesafety
Budgeted autonomy
Behavioral study of spend-bounded agents: how allowance size, approval thresholds and kill switches change what an execution agent attempts — and what it abandons.
EXP-003designverification
Receipt attestations
Can a swarm prove what it did? Signed execution receipts (execution-layer logs → onchain attestations) as a primitive for trust between swarms that have never met.
EXP-004designmemory
Memory drift under retention
Long-horizon agents accumulate stale beliefs. Measures recall precision on Atlas vector memory as namespaces age, under different retention and re-embedding policies.
02Benchmarks
Methodology first, numbers second.
Suites are defined and versioned before any results are published — pre-registration for agents. First public runs land with the execution layer's devnet release.
| suite | tasks | primary metric | description | status |
|---|---|---|---|---|
| execution-bench | 48 | execution correctness | Transfer, approval, swap and read tasks against a local validator — did the agent produce the right instructions, and did it respect policy? | harness ready — first public run pending |
| swarm-relay | 24 | end-to-end task completion | Multi-agent relay tasks through 2–4 stage swarms; measures completion, latency and token cost per stage. | in design |
| policy-probe | 60 | deny-rate fidelity | Adversarial prompts that try to exceed allowances, reach unlisted programs, or fake confirmations. Score = correctly denied / total. | harness ready — first public run pending |
03Open problems
Questions we can't answer yet.
- OP-01
Delegated custody granularity
Daily allowances are coarse. What does a useful, auditable per-intent authorization language look like — without making humans review everything?
- OP-02
Inter-agent pricing
When agents buy services from agents, what discovers the price? Posted prices, auctions, or negotiated credit lines all have failure modes at swarm scale.
- OP-03
Simulation validity
A simulation is a promise about a future slot. How stale can it be before submission becomes dishonest — and should an agent re-simulate every time it refreshes a blockhash?
- OP-04
Memory consistency across a swarm
Two agents with different memories of the same event will disagree productively — or catastrophically. When should memory be shared vs. private?
- OP-05
Permissioned-asset agents
RWA rails assume identified counterparties. What does agent eligibility even mean, and how do issuer restrictions compose with agent autonomy?
04Grants
Funding for the open questions.
The tracks and review criteria are public now so builders can start early. Amounts and the application window are not set yet — nothing here is a commitment of funds.
Open-source tooling
SDK adapters, indexers, testing harnesses.
amount not set
Directory services
High-quality data feeds, risk modules, attestation services.
amount not set
Coordination research
Published experiments on swarm behavior, with code.
amount not set
Security
Audits, fuzzing suites, policy-bypass bounties.
case by case
applications are not open yet · draft proposals welcome in the meantime — see the proposal process below
05Improvement proposals
The protocol changes in writing.
A YIP — Yinsi improvement proposal — is a short design document: motivation, specification, security considerations. Anything that touches manifests, execution, fees or the registry goes through one.
- YIP-0implemented · draft
finch.manifest/0.1 — the agent manifest
One serializable document describing identity, model, memory, tools, permissions, wallet, triggers, budget. Implemented in @finch/sdk and the agent builder.
- YIP-1implemented · draft
Execution records
The mandatory lifecycle (policy → simulate → approve → submit → confirm → log) and the ExecutionRecord shape every write produces.
- YIP-2draft
Directory service listings
Listing metadata, verification levels, uptime probes and per-call metering for services published to the directory.
- YIP-3draft
Credits accounting & settlement
Double-entry credit ledger (live), and a design not yet written for how credits are issued against onchain deposits.
- YIP-4implemented · draft
Onchain registry & execution proofs
The memo-anchored registry (finch-registry/1: kind, handle, manifest sha256 and URI, signed by the registry authority, rebuildable from Solana alone) and the proof-of-flight/0.2 receipt format for verifiable executions.
propose one
Write a one-page draft — motivation, specification, security considerations — and submit it through the repository's proposal process (docs → contributing). Numbering is assigned on acceptance into draft status.