Agent Swarm for Autonomous Liquidity Market Making @solana @MeteoraAG EzZrYUNxh3zJkKCJWMyLkA3Yd21br7bsjMwsBdXqpump https://teapawt.org/t.co/kPG6vIXgmx
Joined July 2026
- Tweets46
- Following21
- Followers147
- Likes18
Pinned Tweet
we've been building this for 2 months straight.
github.com/sentinent-agent/S…
first commit 2 months ago. latest commit 41 minutes ago. public repo the whole time
no stealth, no vaporware. you can read every line and every commit yourself.
verify us on @zauthinc
locked in all day on this.
core engine running the full decision stream end to end: screening, risk evaluation per wallet, routing with expected slippage, position open and close, block sync.
simulated data while we test the surface. real backend events next, and it will read very differently when they land. more refusals, fewer closes.
still on test. back to it tomorrow.
Sentinent neural engine.
What it is, precisely: a model layer that reads each scan cycle and adjusts candidate scores inside a hard clamp, sitting on top of a deterministic layer that makes every actual decision.
Every call is budgeted before it is made, timed out, and schema validated before it is trusted. Any failure returns an empty verdict, and the agent proceeds on its rules as if the model were never there.
The model can be wrong. It cannot be authoritative. That distinction is the entire design.
Terminal demo feed.
The design goal is that every line is something you could verify onchain if it were live: a tx hash, a pool, a realized number. No summaries, no aggregate dashboards hiding the individual decisions.
Wiring to the live backend is in progress. When it lands the same surface will carry real events, including the cycles that produce nothing at all.
Illustrative data throughout.
Sentinent neural engine.
What it is, precisely: a model layer that reads each scan cycle and adjusts candidate scores inside a hard clamp, sitting on top of a deterministic layer that makes every actual decision.
Every call is budgeted before it is made, timed out, and schema validated before it is trusted. Any failure returns an empty verdict, and the agent proceeds on its rules as if the model were never there.
The model can be wrong. It cannot be authoritative. That distinction is the entire design.
Terminal demo feed.
The design goal is that every line is something you could verify onchain if it were live: a tx hash, a pool, a realized number. No summaries, no aggregate dashboards hiding the individual decisions.
Wiring to the live backend is in progress. When it lands the same surface will carry real events, including the cycles that produce nothing at all.
Illustrative data throughout.
shipped today:
- automated rebalance live. up re-centres, down exits. no downward chasing.
- autonomous snipe live. deployer history + narrative filter, hard gates first.
- conversion risk sizing. positions sized against the pessimistic case, not the hopeful one.
- blacklist evaluator running daily. losing pairs get excluded permanently.
- ERC-8004 + x402 deploy path working on testnet.
next:
- terminal wired to real backend events, preview feed retired
- tx classifier decoding deployer activity properly
- mainnet path for 8004 + x402
- holders deploying their own agent
still shipping.
New features unlocked👇
autonomous snipe is live.
how it works:
the swarm has been collecting deployer data on @Pumpfun for weeks. who ships, what they shipped before, how it ended.
when a tracked deployer creates a new coin, the agent doesn't buy on the deploy. it reads the narrative and the lore first. weak story, no entry. if the narrative is strong, it snipes.
hard gates still run before any of that. unsellable token, no buy. it doesn't matter how good the story is.
The order is the point. Payment is verified over x402 before anything touches the chain, and a missing or invalid header returns 402 rather than proceeding. Only then does the agent get registered.
Registration writes name, description, capabilities and a config blob to the ERC-8004 registry, and the receipt returns an agent address, an agent id and a tx hash. The agent stops being a row in our database and becomes an entity with an onchain identity that anyone can resolve.
Body is schema validated before either step. Still on testnet.
shipping two things.
right: conversion risk. before entering, the agent assumes every bin above the active one fully converts to the token at the worst price in its range, and sizes against that number instead of the hopeful one.
left: socket provider. the terminal is being wired to real backend events. stats, leaderboard, feed, config, straight from core.
size for the bad path. show the real feed.
Terminal preview.
The core engine surface: block sync, wallet risk scoring, routing decisions, priority fee adaptation under congestion, and confirmation hashes, rendered as a single stream.
This is the display layer running on synthetic events. The production feed replaces every line with the corresponding backend emission, which is why the scan cycles you have seen posted earlier read differently: fewer trades, more refusals.
Shipping the wiring next.
Terminal preview.
The core engine surface: block sync, wallet risk scoring, routing decisions, priority fee adaptation under congestion, and confirmation hashes, rendered as a single stream.
This is the display layer running on synthetic events. The production feed replaces every line with the corresponding backend emission, which is why the scan cycles you have seen posted earlier read differently: fewer trades, more refusals.
Shipping the wiring next.
New features unlocked👇
autonomous snipe is live.
how it works:
the swarm has been collecting deployer data on @Pumpfun for weeks. who ships, what they shipped before, how it ended.
when a tracked deployer creates a new coin, the agent doesn't buy on the deploy. it reads the narrative and the lore first. weak story, no entry. if the narrative is strong, it snipes.
hard gates still run before any of that. unsellable token, no buy. it doesn't matter how good the story is.
automated rebalance is live.
the part that matters is what it refuses to do. drift up and the range re-centres to follow the move. drift down and it closes.
there is no path in the code that re-centres downward, because that is just averaging into a falling token with extra steps.
panel (a) shows both branches. one of them is not a rebalance, it's an exit.
the swarm, live on sentinent.tech
every node is a pool under watch. the bright cluster in the middle is where the agents are actually looking right now.
right side is the verdict feed. today it reads AVOID down the entire list, because nothing cleared the filters.
watching is most of the job. 👀
big update coming, but the agent swarm never stopped running.
left: telegram. right: the backend logs behind it. same events, same second.
scan completed, 2000 pools sourced, 0 passed the entry filters. it reported the rejection breakdown and went back to waiting.
then the solana stream dropped. auto snipe paused itself, reconnected in one second, resumed.
an agent that refuses everything and pauses when blind is an agent working correctly. 👀
Still debugging.
The brain layer has three properties worth the extra days: calls are budgeted before they are made, responses are schema validated before they are trusted, and any failure returns an empty verdict rather than a guess.
A model that cannot be checked should not be able to move capital. So it doesn't. It advises inside a clamp, and the deterministic path decides.
Shipping when it's clean.
🧑💻
we've been building this for 2 months straight.
github.com/sentinent-agent/S…
first commit 2 months ago. latest commit 41 minutes ago. public repo the whole time
no stealth, no vaporware. you can read every line and every commit yourself.
verify us on @zauthinc
Gud fee tek @MeteoraAG our agent also farm some $CATE
Sentinent retweeted
we've been building this for 2 months straight.
github.com/sentinent-agent/S…
first commit 2 months ago. latest commit 41 minutes ago. public repo the whole time
no stealth, no vaporware. you can read every line and every commit yourself.
verify us on @zauthinc
Today: debugging the testnet path.
x402 payment, ERC8004 agent deploy, identity returned. Now locking in every edge case before mainnet.
A bug found on testnet costs a commit. The same bug on mainnet costs someone's SOL. Worth the extra days. 👀
Lock in $SENTINENT