Agent Farmbeta Join early access

District IX of XNine things

Measured, not claimed

What every step cost, how much plan is left, and whether the work was good, in numbers.

A hand holding a calculator over a folder of bills

An agent always sounds sure, and a dashboard can always look exact. Agent Farm counts instead: what each step cost, read from the transcripts on your machine; how much of every plan is left, as the provider reports it; whether the work passed its check, judged by a cheap model first. Where a figure cannot be known it says so, and it never fills the gap with a guess.

16 Measured

Cost · plan · quality

Every cent counted. Every check scored.

Agents always sound sure. The farm does not take their word for it: it reads what each step cost from the transcripts, keeps every account's window in view, and has the work judged against its criteria by the cheapest model that can tell.

Planevery account · live

Where it wentthe live leg · from its transcript

$1.864.21M raw · ⚖ 0.62M26api·99steps
  • ◈thinking0
  • ▸reply0
  • ⌘local tool0
  • ↩tool result0
  • ⇄MCP tool0
  • ⌗web0
  • ⑂sub-agent0
  • ⚙harness0
  • ·local only0

◈ thinking: billed as output, the dearest token there is.

  • ⚑ editor is at 4.5× its normal today: $4.10 against a $0.92 median
  • $0.00 OpenCode on a free model: a real zero
  • tok Codex on a ChatGPT plan: tokens, and why there is no price; never one made up

Judge laddera step's check · cheap first

Checkshot-3.png“the beam is visible through the rain”

    Evals · after a model change

    cases passed47 / 50−1
    tests passed100%=
    Insights2 findings1 rule suggested
    IX·1

    Live plan usage

    How full each account's five-hour window and week are, and when they reset, in the status bar and on the Dashboard. Every engine's limit apart.

    Every account's plan, live

    What it solves

    The limit of a subscription plan arrives in the middle of a turn, without warning. Typing /usage in Claude Code tells you where you stand, if you remember to, and usually after the long turn rather than before it.

    How it works · pick a step

    Beside it, a lap counter of your own: Reset Token Lap counts the whole fleet's tokens, cost and messages from that moment, as ⏱ 84.2K in the status bar, like a car's trip meter. How the percentage, the time left and the reset are drawn is yours to change (quotaStyle.*).

    Settings · 5
    SettingDefaultWhat it does
    agentFarm.planUsageWarnPercent80%The status bar turns yellow.
    agentFarm.planUsageDangerPercent95%The status bar turns red.
    agentFarm.planUsageActiveRefreshSeconds60 sHow often the plan is read while an agent works. Never under 30 s.
    agentFarm.planUsageRefreshMinutes5 minHow often an idle farm reads it. 0 = only when you press ⟳.
    agentFarm.planUsageStatusBaronThe status bar indicator. The card stays either way.
    IX·1The status bar1 / 5
    IX·2

    Where it went

    Every turn split into nine kinds of step, from the model's thinking, billed as output, to local tools that cost nothing to run but make the next call bigger. Every figure is shown raw and weighted.

    A live leg, counted by kind

    What it solves

    "Where did my limit go?" had no answer. Reading a file and nine seconds of the model's reasoning drew the same row, though one runs here for free and the other is billed at the output rate. And a raw token total is mostly cache reads, the cheapest thing in it, so it does not describe the wall it seems to.

    How it works · pick a step

    The rates are the provider's, so they are settings, never a release. Changing the weights re-reads every transcript at once, and a table that cannot be used falls back to the shipped one rather than quietly zeroing the gauge.

    Settings · 4
    SettingDefaultWhat it does
    agentFarm.quotaWeightModemodel-awaremodel-aware: Opus 5, Sonnet 1, Haiku ⅓. model-flat: weigh only the kinds of token. raw: ⚖ is the raw total.
    agentFarm.quotaWeightsthe tableThe weights per model, against one Sonnet input token. Apart from the prices, so a billing fix never moves the gauge.
    agentFarm.windowWeightedTokenLimit0 (off)Your five-hour ceiling in weighted tokens, for a percentage.
    agentFarm.pricingthe rate cardUSD per million tokens per model pattern; the first match wins.
    IX·2Nine kinds1 / 5
    IX·3

    Budgets

    A daily budget per agent or for the whole fleet, a cap per workflow and per mission, a monthly media budget. A warning at 80%, an error past 100%.

    Budgets filling, live

    What it solves

    An agent left to run overnight can spend what was meant for a week, and a warning in the morning is too late. A ceiling you can walk through is not a ceiling.

    How it works · pick a step

    On a subscription the dollars are the API-equivalent value of the work, not a bill: "builder should not use more than $25 of value today". Seven places read the budgets, and all seven read them through one door, so the switch that lifts them is obeyed everywhere or nowhere.

    Settings · 4
    SettingDefaultWhat it does
    agentFarm.budgets{}Daily USD per agent; "*" = the fleet.
    agentFarm.budgetsUnlimitedoffSuspends every daily budget without losing the figures.
    agentFarm.workflowSpendCap0 (none)A run's hard cap; a workflow or a single run can set its own.
    agentFarm.mediaBudgetMonthUSD0 (off)A warning when the last 30 days' media cost passes it.
    IX·3Per agent, per fleet1 / 5
    IX·4

    Cheaper, not stopped

    As a window fills, automatic jobs wait for it to loosen, or move to a cheaper model. Sessions you opened by hand are left alone.

    Windows past the line, live

    What it solves

    An overnight factory eats the morning's quota, and a job that hits the wall halfway through is a job done for nothing. Cancelling it is not the only answer.

    How it works · pick a step

    Settings · 5
    SettingDefaultWhat it does
    agentFarm.quotaGuardPercent0 (off)Automatic jobs wait at this fullness of the five-hour window.
    agentFarm.quotaDowngradePercent0 (off)Past this fullness, the farm's own jobs run on the cheaper model.
    agentFarm.quotaDowngradeModelhaikuThe cheaper model.
    agentFarm.quotaHandoverPct90%The handover line; 0 turns the hold off.
    agentFarm.quotaBrakeBasiswhicheverWhich figure the guard and the downgrade read: raw, weighted, or whichever is further along.
    IX·4Wait1 / 5
    IX·5

    Cost alarms

    A day that costs far more than an agent's usual, or a week that jumps, raises a flag. The usual is the median of its real days, not a guess.

    A cost alarm, live

    What it solves

    A budget has a blind spot: a day that costs ten times the usual but stays under the cap passes in silence. And cost creeps, a little more every week, with no single day worth a warning.

    How it works · pick a step

    A click on the flag opens that agent's own cost view, where the evidence is. Dismiss it for a day, a week or for good, and tomorrow's "3.9×" stays the same dismissed finding as today's "4.2×".

    Settings · 2
    SettingDefaultWhat it does
    agentFarm.costAnomalyMinUsd$1Days under this are not judged. 0 = off.
    agentFarm.trendAlertFactor2×This week against last. 0 = off.
    IX·5Fourteen days1 / 5
    IX·6

    Honest numbers

    Costs come from your own transcripts. A free model's zero is a real zero; a missing figure is shown as missing.

    What it solves

    A figure that looks exact and is not is worse than none: $0.00 beside work that was billed, a session that "ran 36 hours" because it was picked up again the next morning, an hourly cost divided by a night's sleep.

    How it works · pick a step

    An interrupted or cancelled turn can leave lines that report zero usage. They are not read as an empty context: the real figure stays.

    Settings · 2
    SettingDefaultWhat it does
    agentFarm.sessionBreakMin30 minA pause longer than this between turns is not counted as work. 0 = count everything.
    agentFarm.agentRowCostUnittokensWhat each agent row in the Agents view shows for today: tokens or cost.
    IX·6From your transcripts1 / 5
    IX·7

    Judge ladder

    A check is judged by a cheap model first. Only a fail climbs to a stronger one, and a pass there overturns it, so no retry is paid for nothing.

    The ladder, judging

    What it solves

    Checking every step with a strong model costs more than many of the steps. Checking with a cheap one alone fails good work, and every wrong FAIL buys a retry nobody needed.

    How it works · pick a step

    Three cost presets set a run's ladder and cap together: Economy judges on haiku with a $5 run cap, Balanced is the shipped behaviour, Quality climbs from sonnet to opus with no cap.

    Settings · 3
    SettingDefaultWhat it does
    agentFarm.judgeArgshaikuThe judge's model. Keep it cheap.
    agentFarm.judgeLadderemptyModels cheapest first, such as haiku,sonnet. Empty: one judge, no climb.
    agentFarm.judgeCompareautoShows the judge what an edit started from. never: the result alone.
    IX·7Cheap first1 / 5
    IX·8

    Evals

    Saved cases replayed after a change of model or instructions: accuracy, tests passed and what got worse, on a scorecard, not on how sure an agent sounds.

    An evals scorecard, live

    What it solves

    After a change of model or of an agent's instructions, the only evidence left is how sure the agent sounds. A change that breaks yesterday's good answers goes unnoticed until someone trips over it.

    How it works · pick a step

    A quality average that drops by 0.8 in a week, or sits under 3 for two, raises a warning. Cheap and bad is a loss, so cost and quality sit side by side.

    Settings · 4
    SettingDefaultWhat it does
    agentFarm.evalRunArgs--model haikuThe model eval cases are replayed on. Empty: the agent's own.
    agentFarm.qualitySampleoffScore recent sessions: daily or weekly.
    agentFarm.qualitySamplesPerAgent2Sessions scored per agent each time.
    agentFarm.pulseoffThe pulse check: nightly or weekly.
    IX·8A case1 / 5
    IX·9

    Insights

    Repeated corrections turn into suggested rules. A finding is fixed, handed to its agent or dismissed, and whether the fix held is tracked.

    Findings and a suggested rule, live

    What it solves

    A finding that is only a snapshot gets fixed, comes back, and nobody knows whether the fix held. And the same correction typed to an agent every day never becomes the rule it should be.

    How it works · pick a step

    The Living page exists because of a measured failure: a daily cycle stopped for 16 days while its reminder ran correctly every morning and nobody read it. No error, nothing red, and the system was dead.

    Settings · 3
    SettingDefaultWhat it does
    agentFarm.insightAutoDelegate[]Categories handed to their agent as soon as they appear; at most one task per problem a day.
    agentFarm.suggestScanoffScan for rule suggestions every week.
    agentFarm.livingChecks[]Your own checks on the Living page: a diary file, a folder that should be empty.
    IX·9One list1 / 5

    Questions

    Is the cost a bill?

    On a subscription, no: it is what the same work would cost at the provider's API rates, the value you got from the plan. On an API key it is close to the bill, at your own rate card. The plan figures, how full a window is, are the provider's own.

    Where do the plan percentages come from?

    From the provider: for Claude, the figures /usage prints, read with the sign-in Claude Code already keeps; for Codex, its session file; for Antigravity, agy -p /usage. Everything else on this page is counted from the transcripts on your machine.

    Will Agent Farm stop or cheapen my work by itself?

    Only where you said so. The guard, the downgrade and the run cap start off, and budgets start empty. Sessions you open by hand are never moved to a cheaper model.

    Do the checks cost money?

    A little, and they are counted. The judge is a cheap model, a lean check on haiku was measured at about a fifth of a cent, and its cost is added to the step it checked. An eval run or an A/B test asks before it spends.

    What happens when a provider changes its prices?

    You change the rate in settings or in the Model pricing editor; no update is needed. The new rate applies from then on, and what was spent before keeps the price it had.

    Why is there no price beside Codex?

    Codex on a ChatGPT plan reports tokens, not a price. Agent Farm shows the tokens and says why there is no money beside them, rather than estimate one from Anthropic's rates.