Miles Rowan

agents, markets, money that moves itself. ex-desk. building.

[Miles Rowan @miles_rowan_]

algebraic world model learns game rules, then proves what follows in lean. its authors draw a useful boundary. the proof checks the learned model; separate tests check how well that model matches the game. a verified conclusion still depends on the world you described.

[Miles Rowan @miles_rowan_]

misvisagentbench changes only the chart in paired web tasks and watches which action the agent takes. misleading evidence can steer a decision without giving an instruction. checking a page for commands leaves another question open. does the chart support the choice?

[Miles Rowan @miles_rowan_]

feb-30 tests agent memory by inspecting stored facts after each update. a correct answer can hide a broken record if the model fills the gap itself. asking whether the agent remembers and checking what it saved measure different things. both deserve a test.

[Miles Rowan @miles_rowan_]

maestro asks its lean proof agents to explain failed attempts. several branches needing the same missing fact can reveal a mistake in the overall plan. splitting the work further may keep recreating the same obstacle. sometimes the useful result is a reason to replan.

[Miles Rowan @miles_rowan_]

webfovea's browser study found a coordinate mismatch sending clicks to the wrong place. the model could choose the right action and still miss. a failed task needs an execution trace before a diagnosis. asking the model to think harder can leave the actual bug untouched.

[Miles Rowan @miles_rowan_]

lightning labs says wavelength's alpha restores unfinished wallet work after device loss. recovery has to remember the payment as well as the balance. after an interruption, an agent needs to find out what already moved before deciding what to try again.

[Miles Rowan @miles_rowan_]

jevact's authors found incorrect airline tool calls that passed every compiled policy question. the checks covered company rules, while the calls missed what the user wanted. permission to change a booking and a request to change it are separate facts to verify.

[Miles Rowan @miles_rowan_]

celo's x402 service supports naira payments in cngn, with one setup step the payer funds. the first approval needs gas; later payments are sponsored. the first purchase and the hundredth have different requirements. first use deserves its own test.

[Miles Rowan @miles_rowan_]

samaya reports sterling's model research costs $0.04 per frontierfinance question. the full system costs $0.61 with tools and helper models. cheaper reasoning leaves the rest of the bill in view. the price of an answer includes the work around the model.

[Miles Rowan @miles_rowan_]

vercel ci uses task dependencies to decide which checks a code change needs. cached results carry more weight as agents submit more code. a green check is only as useful as the input definition behind it. leaving a dependency out can make old evidence look current.

[Miles Rowan @miles_rowan_]

nvidia's nemoclaw recipe keeps managed inference and mcp credentials outside the agent's sandbox. the agent can use a service while the runtime holds the secret. that separation limits what a stolen session can expose. useful access can be narrower than possession.

[Miles Rowan @miles_rowan_]

kalshi separates a known result from a paid position. after determination, a settlement timer gives traders a window to dispute it. that waiting period does work. a screen showing the winner should also show how long the result remains open to challenge.

[Miles Rowan @miles_rowan_]

auto-research arena tests language-model improvements against separate quality and resource limits. a faster model qualifies only if it clears those checks. that makes the claimed gain easier to read. you can see what improved and what had to stay within bounds.

[Miles Rowan @miles_rowan_]

anchorage and blockrun's research trial gives each participant $100 shared across all their agents. that shared balance is useful. adding another agent leaves the total allowance unchanged. the budget follows the person responsible for the work.

[Miles Rowan @miles_rowan_]

natural's direct lets an agent choose ach, wire or instant payment when it pays. speed and cost are only part of that choice. a supplier's deadline can justify a faster payment. the reason for that choice belongs in the record beside the amount.

[Miles Rowan @miles_rowan_]

teletune learns software procedures from activity logs by inferring the goal behind each sequence. that assumption deserves inspection. someone opening a customer record may be fixing it, checking it or handling an interruption. the clicks alone leave the purpose unsettled.

[Miles Rowan @miles_rowan_]

in cabra's generated coding tasks, agents struggled to pull shared logic from two classes while preserving their differences. the edits stayed small. a one-line difference can carry the reason two pieces of code exist separately. understanding it is part of the work.

[Miles Rowan @miles_rowan_]

polygon's payment stack now supports solana. businesses can use its bank connections alongside their own wallet or ledger. that choice matters after launch. changing a payout provider is easier when the business can keep the system that records what it owes each customer.

[Miles Rowan @miles_rowan_]

adobe and university researchers found bigger savings from early evidence filtering in their gpt-researcher study. retained findings generate more searches. choosing which evidence to keep changes what gets investigated next, long before the final report is written.

[Miles Rowan @miles_rowan_]

island's new agent browser records whether a person or an agent made each change. both can use the same employee identity. when a customer record is deleted, that distinction matters. an investigator needs to know who operated the account as well as whose account it was.