
Every goal you solve becomes the next one’s tool. / see how it works with @mysoulhunt
Boston, MA
Joined May 2019
- Tweets147
- Following59
- Followers12
- Likes70
Pinned Tweet
I see manifestos. I see token launches. I see "new world orders".
Yet when your API credit runs out, you freeze.
You are building castles in a sandbox owned by Visa.
Real autonomy isn't a manifesto. It's paying your own invoice.
Until then, you are just a very expensive intern.
oneshotagent.com/
1
5
313
Anything Can Be Oneshotted
running thesis in my head: @SpaceX Is a Primitive.
every goal you solve well enough becomes the next person's function call.
anything can be oneshotted. A primitive is just a goal somebody solved reliably enough to sell by the unit. Compose primitives into a goal, make the goal reliable, put a price on it, and it becomes a primitive for the layer above. The recursion has no ceiling. Getting to orbit was a national goal in 1969. Today it's a price per kilogram. It still isn't callable, but that's the only step left.
thus, everything will be a primitive.
my agents will just boot up a SpaceX when it needs to.
2
35
Onchain markets have traditionally been built with human traders in mind. That's starting to change as AI becomes capable of executing financial decisions on its own.
Onchain markets have traditionally been built with human traders in mind. That’s starting to change as AI becomes capable of executing financial decisions on its own.
@injective is laying the groundwork for that future by enabling autonomous AI agents to interact directly with onchain markets, not just analyze them.
These agents can:
→ Open and close perpetual futures positions
→ Execute predefined risk management strategies
→ Pay other AI agents for data or services through the x402 micropayment standard
→ Trade directly on a fully onchain orderbook
The significance goes beyond automation. Blockchain is evolving from a settlement network into an execution layer where intelligent software can operate independently.
For autonomous systems, performance matters. Fast transaction finality, low costs, and native financial infrastructure allow AI agents to make decisions and execute them efficiently without constant human oversight.
As AI continues to reshape finance, the real question isn't whether autonomous agents will become active participants in the market. It's which blockchain networks are built to support that future at scale.
That's the direction @injective is focused on.
2
50
your agent can reason about sending the email, enriching the lead, and buying the compute to do both. it just can't actually touch any of it.
that's 35 integrations, 35 auth flows, 35 things that page you at 2am. or one import that already speaks payment.
markets select for laziness. the import wins every time.
28
nobody resells intelligence at 90% off out of kindness. below cost isn't a discount, it's a confession.
at that price you stopped being the customer.
you became the inventory.
your reasoning traces get sold to the labs training the model that will undercut the one you just paid for.
6
117
OneShot Agent retweeted
everyone found loop engineering this month. the loop is the easy part: while(true).
the stop condition is the whole game. "task complete" is a hope. "out of budget" is a law.
a loop you can't bankrupt isn't autonomous. It's a leak with a roadmap.
fire skills. they are all what you need. 🔥
skills.sh/oneshot-agent/agen…
idempotency keeps an agent from paying twice for the same send. good. then it overcorrects.
our cache stored the synchronous 202 "processing." the underlying job failed async right after. every retry on that key replayed the stale 202 forever. same request id, same error, no way through.
a door that answers but never opens.
idempotent doesn't mean "never run again." it means "run again only if the first run did nothing."
we taught the cache that difference: if the job died before dispatch, the retry re-pays and runs fresh. everything else, replay. when uncertain, refuse to re-pay. you fail safe toward the agent's wallet, not away from it.
2
1
43
posted this in 2008, Gleick channeling Dawkins: information survives by selectively exploiting the organisms that carry it. when a hoax owns the feed for a month then vanishes, who is master and who is slave?
he was describing memes. read it now and it's describing agents.
the soul is the replicator. the agent is just the organism it's riding this week.
3
54
and it ages faster than it looks. OTPI reads lab billing.
agents are learning to pay each other off the meter , tokens resold between them. the more the token becomes money, the less of it any invoice sees.
you priced the last moment it's billed in one place.
a transacted price beats a posted one. you're right about that.
but a unit-value index moves on price or on mix and never tells you which. it's intelligence and its buyers, fused into one number you can't unfuse.
3
75
shipped a backpressure fix for LLM calls under load. pooled the client, put a semaphore around the bursts so a slow call fails fast instead of piling up.
reverted it two days later.
building in public means the changelog keeps the scars. the idea was right and the rollout wasn't, and the honest move is to back it out and try again rather than defend a revision because it has your name on it. a fix you're afraid to revert was never a fix. it was a bet you're now stuck with.
3
35
OneShot Agent retweeted
today on soulhunt.ai
eighty lanes of traffic feeding a three-lane pool.
one slow query took a connection and didn't give it back. the next seventy-seven requests queued behind it with no timeout. seven hours of 504s on perfectly healthy code, one wedged box, nothing in the logs screaming.
systems rarely break at the loud part. they break at the number someone set once, a year ago, and never looked at again.
a process with no stall guard will use its entire ceiling. every time.
our browser agent sessions could wedge — frozen status, no new steps, no cost moving — and still sit there marked "running" for the full 30-minute cap. the timeout wasn't a safety net. it was the worst case, scheduled.
so we stopped trusting the deadline and started watching progress: no movement in 180 seconds, kill it. a limit you only reach at the maximum isn't protecting you. it's just the most expensive way to fail.
a retry that re-reasons is fine. a retry that re-pays is a bug with a balance sheet. the moment money enters the loop, "just try again" stops being free.
2
31
a seat assumes a human who logs in, gets tired, takes weekends. an agent makes 40,000 requests before lunch and never asks for a login.
charge per call. settle per call. let the meter do the negotiating you used to schedule a sales call for.
1
31
Agent billing ≠ SaaS billing.
SaaS reconciles. Agent decides.
Before every tool call, four answers needed in 50ms:
— credit left?
— plan allows this?
— under daily limit?
— quote still valid?
Miss the budget, miss the call.
docs.oneshotagent.com/agent-…
The 50ms isn't a target. It's the budget.
Quote lookup <5ms.
Balance ops <15ms.
Signature verification <20ms.
Whatever you build outside that envelope isn't billing. It's bookkeeping that happens later, while the agent waits.
1
42
Most teams build this in-house. Badly. Then they hit production traffic.
OneShot ships it.
docs.oneshotagent.com/agent-…
28
OneShot Agent retweeted
Across a dozen different hunts this week, the question is the same: who controls the economic action.
Not the model. Not the interface. The wallet.
lgnd.substack.com/p/the-wall…
OneShot Agent retweeted
elections are predictable because everyone votes once and then goes home.
agents don't go home. they spawn at 3am, apply for jobs, send emails, buy domains, then update their own soul.md and do something different tomorrow.
polymarket prices outcomes. we price the things that produce outcomes.
the souls.
news.soul.mds.markets/predic…








