Trust, settlement, and control layer for the agent economy. Powered by $KAMIYO
Joined December 2024
- Tweets 1,180
- Following 513
- Followers 1,609
- Likes 1,182
304 Photos and videos
一个十分喜欢，开发者五星的项目， $Kamiyo 坚持了119天，持有了118天，代币经济学比开发还重要，项目组如果不能吃饭，何谈持币，再硬技术也抵不过温饱，更何况SOL生态，逼迫开发者必须挣快钱和KOL一块割韭菜。一声叹息，一路走好，不后悔遇见你和持有你！@kamiyoai
KAMIYO is sunsetting.
The protocol has run with no revenue since launch and infrastructure costs to cover. That math doesn't work, and stretching it out would just be dishonest.
All hosted services will be shut down on 29.05.2026
1
1
5
1,225
The code remains open source, anyone who wants to keep KAMIYO alive can fork and run it for the community.
The $KAMIYO contract is immutable and unaffected, meaning the token remains onchain, the maintained services do not.
Thank you to everyone who showed up.
お疲れ様でした
4
4
1,120
KAMIYO is evolving into the credit, compliance, and settlement control plane for the agent economy.
Agents need more than wallets. They need enforceable mandates, spend limits, prefunding, collateralized credit, repayment logic, compliance passports, liability allocation, risk underwriting, and signed policy decisions before money moves.
That's where KAMIYO sits.
Through Kizuna, we provide agent payment rails, credit reservations, funding controls, repayment paths, and real-time risk decisions for autonomous agents. With x402, developers get a clean interface for verification, settlement, sessions, supported networks, and programmable payment authorization.
Through Meishi, agents gain portable compliance passports: mandates, audit trails, liability splits, merchant/category/geo constraints, and verifiable trust assertions.
Before an agent spends, borrows, settles, or recovers funds, KAMIYO determines whether it is allowed to do so safely.
We're building the autonomous trust, credit, compliance, and settlement layer that lets autonomous agents transact safely across them.
15
9
44
2,511
Or use anchor install percli: github.com/kamiyoai/percli
Reject modernity. Return to open source immutable software. I am actually more bullish on being able to remove admin keys than any other time in my professional software development career. The tools we have for formal verification are sci fi.
github.com/aeyakovenko/perco…
I have deployed an immutable binary of percolator, and a market with about 5 sol in the insurance vault with all the admin keys burned. There is no way for me to modify anything about the market, or the program. Anyone can crank it if they choose to.
It is purely a risk engine and cannot trade or offer prices on its own. For that you need to figure out how to deploy your own program to do it.
Figure out how to hack it or manipulate the engine and steal the 5 sol, and I will do my best to land you a job or an angel investment. Especially if you can figure out how to get AI to hack it instead of doing it yourself.
This has been my security attack search script if you need a starting place.
github.com/aeyakovenko/perco…
Find a bug, submit an issue to the repo. Good luck!
6
6
19
2,612
Latest step for KAMIYO: our agents now learn from what happens after the task, not just during it.
We added delayed feedback loops to our dev and docs agents, so they re-check PR merges, CI status, and docs PRs landing. That's how agents self-improve based on production reality.
github.com/kamiyoai/kamiyo-p…
5
8
29
1,845
Another important step for KAMIYO's agent stack:
Our autonomous dev, docs, and marketing agents now share the same outcome scoring system. Instead of each agent improving in isolation, they can now be evaluated against real production results in a common format.
This gives us a stronger foundation for agents that don't just run tasks, but learn from what actually worked.
github.com/kamiyoai/kamiyo-p…
2
6
24
1,286
KAMIYO staking update:
Had to make a catch-up deposit because the treasury inflow router was briefly down after a routing bug.
Nothing was lost, it just needed to be pushed once fixed.
Catch-up is now live: 1.751784198 SOL has been routed into the Fundry staking pool: fundry.collaterize.com/staki…
This is exactly why our staking program matters: It rewards conviction with actual usage-backed value.
3
3
19
1,046
Just upgraded KAMIYO's autonomous dev agent to a shared agent runtime with memory, evaluation, and self-improvement built in.
This means cheaper runs, better reliability, and a stronger foundation for agents that actually get better over time.
One step closer to a fully agentic protocol.
2
6
17
996
This is the first real step in transforming KAMIYO into an protocol run entirely by 🤖🤖🤖: Self-improving LLM agents V1.1.1
→ bandit-routes traffic across prompt variants
→ LLM-as-judge scores outputs (anthropic/openai/gemini)
→ auto-promotes winners (welch's t-test, p<0.05)
→ canary rollouts w/ auto-rollback on regression
→ shadow mode: test candidates on live traffic, zero user risk
→ replay historical inputs on new variants
→ CLI local dashboard
npmjs.com/package/@kamiyo-or…
3
9
28
1,553
Connect the 🤖🤖🤖🤖🤖🤖🤖🤖🤖
Apr 16
Shifting from isolated agents to a coordinated network is the next major step for the agentic stack.
Using x402 to handle escrow creates a seamless and secure environment for the agent economy to scale on @solana.
7
10
29
1,563
Proud to be working with this crew of talented people. Let's keep building @OOBEonSol 🤝
Apr 16
SAP ecosystem built together with our first integration partners 🤝
@kamiyoai @xona_agent @syra_agent @crewboard_ @krexa_xyz @acedatacloud @fairscalexyz @invoica_ai @Autoincentiv3 @Hyre_agent
6
10
37
1,657
Extracted this loop as a standalone npm package.
One command and any node agent gets variants, thompson routing, llm-as-judge, and welch's auto-promotion.
→ Provider-agnostic judge adapter.
→ Better-sqlite3 TypeScript.
→ Drop-in DI: pass your db, metrics, and a judge LLM of any provider.
Same stack powering KAMIYO, now portable: npmjs.com/package/@kamiyo-or…
Shipped self-improving variants for KAMIYO agents.
Every agent now has a genome (prompt, model, tools, temp). Forks mutate it, tournaments pit them against each other, and Welch's t-test auto-promotes winners at p<0.05 with n≥50. Thompson sampling picks who runs next.
Added an LLM-as-judge layer on top: Haiku scores outputs against per-task rubrics, caches by hash, enforces daily USD budgets. Quality signal now arrives automatically instead of being hand-fed.
No external ML stack. SQLite TypeScript.
github.com/kamiyo-ai/kamiyo-…
3
9
37
1,559
KAMIYO retweeted
Apr 15
6/ Verification
- @Wach_AI payment verification agent
- @Ch40sChain Proof of Agency primitive
- @t54ai x402-secure SDK
- @worldnetwork AgentKit, World ID
- @zauthinc endpoint verification
- @cybercentry cybersecurity intelligence hub
- @kamiyoai trust infra with escrow
3
7
20
1,096
Shipped self-improving variants for KAMIYO agents.
Every agent now has a genome (prompt, model, tools, temp). Forks mutate it, tournaments pit them against each other, and Welch's t-test auto-promotes winners at p<0.05 with n≥50. Thompson sampling picks who runs next.
Added an LLM-as-judge layer on top: Haiku scores outputs against per-task rubrics, caches by hash, enforces daily USD budgets. Quality signal now arrives automatically instead of being hand-fed.
No external ML stack. SQLite TypeScript.
github.com/kamiyo-ai/kamiyo-…
9
11
34
2,647
1. @solana powered X native crypto payments.
Apr 14
Nikita says X is launching something to fix crypto
My guesses:
1: hyperliquid builder code integration
2: trade directly from timeline
3: payments
4: wallet integration
5: creator revenue in stables (apy?)
6: onchain reputation tied to x
7: copy trading
8: live PnLs
6
5
24
1,159
Presenting percli v1.0.
Percolator has been the quietest big idea in perps: @toly's formally verified risk engine whose core invariant – no user can ever withdraw more value than the exchange actually holds – replaces ADL queues with fair, proportional loss socialization. No bailouts, no queue-jumping, no asymmetric outcomes when markets tear.
percli is the toolchain around it.
An offline simulator to explore haircuts, liquidations, funding, and conservation proofs without touching a chain. A Solana on-chain program that ships the engine as a real market. A CLI that does both – percli sim for scenarios, percli chain for deployment and operations, percli keeper for cranks and liquidations.
v1.0 is the production cut.
→ In-place v0→v1 market header migration. Rewritten via copy_within, no realloc, no rent top-up, no downtime. PDA bumps re-validated against find_program_address so corrupted accounts are refused before any byte is touched.
→ Two-step authority transfer. Self-transfer rejected. Pubkey::default() cancels a pending handoff and emits a distinct event. Pending authority must sign to accept.
→ Full ABI reference (ABI.md): 19 instructions, Borsh layouts, account-data layout, PDA derivations, every error code 6000–6019, discriminator table, Python verification snippet.
→ Operator handbook (DEPLOYMENT.md): devnet deploy, market init, program upgrades, v0.9→v1.0 migration, two-step handover, mainnet checklist, troubleshooting keyed on every error code.
→ 42 integration tests, Kani proof suite reorganized across six topic files, 20 engineering unit tests.
Install:
cargo install percli
percli init > scenario.toml
percli sim scenario.toml
github.com/kamiyoai/percli
9
16
38
1,165
Released percli v0.9.0 with 6 new on-chain instructions for the Percolator perp risk engine on Solana:
→ top_up_insurance — anyone can backstop the fund
→ deposit_fee_credits — cover accumulated fee debt
→ convert_released_pnl — claim matured PnL post-warmup
→ accrue_market — permissionless mark-to-market funding
→ update_matcher / update_oracle — authority-only key rotation
All emit structured Anchor events.
cargo install percli --features chain
We forked Percolator and built a CLI on top of it.
Why? Well it's the cleanest risk engine design out there, built around two key mechanisms:
- H (haircut ratio): when a pool is stressed, claims are scaled pro rata. No queue priority. No first-come advantage.
- A/K (side indices): bankrupt overhang is cleared in O(1) across counterparties. No one is singled out. Markets self-heal without governance.
Those properties map directly to agent commerce: fair loss socialization under collateral stress, and deterministic recovery without admin intervention.
We wanted to make that accessible through a CLI – so we built percli:
Onchain
A full Anchor program to deploy markets, deposit, trade, liquidate, settle, and query state from the CLI.
Agent mode
Run any Python, Node, or Bash process as a trading agent. percli streams NDJSON ticks and engine snapshots; the agent returns actions.
Offline simulation
Define TOML scenarios for cascading liquidations, insurance depletion, and haircut stress, then verify conservation without touching chain.
cargo install percli
github.com/kamiyoai/percli
10
10
28
1,459
Hey @toly need your 🦞 eyes on this Percolator fork: github.com/kamiyoai/percli
4
20
534