Session budgets — the fuel-up-and-go pattern

Status: Candidate — awaiting founder verification. Why this page exists: The single most important UX pattern for keeping an LLM-powered product from going broke.

TL;DR

A session budget is a hard dollar cap on a single agent run. You pre-allocate the budget, the agent works inside that cap, the session ends when the cap is hit or the work is done. No surprises. No runaway costs. Fuel up, go, refuel later.

The deeper read

Open-ended agent runs are the most common way LLM-powered products bleed money. The agent loops, retries, decides to "do more research," and three hours later you're staring at a $40 bill for a task that should have cost $2.

The fix is bounded sessions. The user (or the founder) says: this session has $X. Do as much as you can. Tell me what's done. Stop.

The pattern is borrowed from how indie devs already work — pre-paid mobile plans, hourly studio bookings, prepaid taxi rides. You decide the budget before the work starts. You can't accidentally spend $40.

The session lifecycle

1. ALLOCATE
   Founder opens the session launcher.
   Sets a dollar budget: $5, $10, $20, $50, $100.
   Confirms with "FUND" type-to-confirm.

2. PLAN
   Foundation Team's Estimator reads the requested work and breaks
   it into stages with per-stage cost estimates.
   "$2.50 for Cartographer scan + $1.20 for Archivist + $0.80 for
    Strategist. ~$4.50 total. Fits in your $5 budget."

3. EXECUTE
   Agents run inside the budget envelope.
   Quartermaster tracks spend in real time.
   At 80% of budget, the system pauses and asks: continue, stop,
   or top up?

4. REPORT
   Session ends. Foundation Report written to the substrate.
   Quartermaster logs the final spend.
   Audit log records what was done and what it cost.

5. RESUME LATER
   Next session, the work picks up from the last checkpoint.
   No state lost. No re-discovery cost.

Why hard caps matter more than soft warnings

Soft warnings ("you've spent $4 of your $5 budget") don't actually stop spend. They notify you, then the agent keeps going. By the time you read the notification, the work is done — and the bill is real.

Hard caps stop spend. The Quartermaster agent refuses to launch the next tool call if it would push the session over budget. The session simply ends.

The tradeoff: you might end mid-task. Apiary handles that with checkpoints — every Foundation Team agent writes its progress to disk continuously, so resuming means reading the file, not redoing the work.

Budget tiers we found useful

$5    Single foundation scan, no synthesis.
$10   Scan + light synthesis. ~30 minutes of agent wall-clock time.
$20   Full Foundation Team pass. ~2 hours wall-clock.
$50   Synthesis + a small generation pass (~2K examples).
$100  Bigger generation pass (~10K examples).
$500  Fine-tune budget (covers GPU rental + data prep).

The amounts are deliberately small. Apiary is designed for indie founders, not enterprise teams. A $5 session is a real unit.

How the UI surfaces this

The founder console (Tab 7 — Foundation Team) has a session launcher:

Pick a budget from the tier list.
Pick which agents to run.
Type "FUND" to confirm.
Watch the live spend bar tick up as the agents run.
See the report when it's done.

The pattern is fuel up and go. No subscriptions. No surprise charges. No "wait, why was this month $400?"

Source quotes

"Session A — $20 budget. Generator produces about 2,000 synthetic examples. Curator filters them down. Validator samples 100 for your eyeball. Session ends. Foundation DB updated. You review. Next session: resume from where Session A left off. You're never on the hook for more than the session's budget."