A multi-agent desk without burning your API budget

A multi-agent demo is easy to film and expensive to leave running. One agent plans, one searches, one writes, and a fourth “orchestrator” summarizes the other three. By Thursday the API dashboard looks like a leak. Bix is a calmer shape: one desk, three pals, one snack jar.

This is the operating manual for that desk. It assumes you have already read why prepaid beats leftover seats. If you have not, read that first, then come back and put Atlas to work.

Why three agents, not one blob

A single generic assistant will plan, research, and draft in the same tone. That is convenient and mushy. Bix splits the jobs:

• Atlas — planner pal. Tiny next steps. Cheerful, not bossy.
• Scout — curious pal. What we know, what we do not, what to peek at next.
• Quill — word pal. One strong version, then variants if they earn their snacks.

Meet them properly in Meet Atlas, Scout, and Quill. The important budget fact: they share credits. Switching pals is not opening a second vendor.

The burn patterns that kill multi-agent projects

Parallel fan-out

You send the same prompt to three models and pick the prettiest. That is a bake-off, not a desk. On Bix, pick the pal whose job matches the ask. Atlas does not need to rewrite the FAQ. Quill does not need to invent a research plan.

Context stuffing

You paste the entire brief into every turn. Tokens are not a personality. Scout can carry assumptions in a short list. Quill needs the audience and a sample, not your seed deck.

Orchestrator theater

An extra agent that only says “I will now ask the writer” is a snack tax. Bix lets *you* be the thin orchestrator. You are cheaper than a fourth model call.

Invisible meters

API keys hide in `.env` files. Prepaid jars sit on the desk. If you cannot see the bite, you will not stop. That is why Bix hard-stops at zero and why the calculator exists.

A desk workflow that stays cheap

Use this loop. It is boring. Boring is solvent.

1. Name the job in one sentence. If you cannot, ask Atlas *only* for a five-line plan.
2. Scout the gaps. Ask Scout what is unknown. Do not let Scout write the launch post.
3. Draft once with Quill. One strong version. A second version only if a real reader would notice.
4. Check the jar. If the next ask is curiosity, not shipping, stop.
5. Refill on purpose. Packs are $10 / $25 / $50. Stripe, not “we’ll reconcile tokens later.”

That loop is also the weekend launch checklist, compressed.

Mapping work to pals without a router model

You do not need an LLM to route to an LLM. Use the URL. Bix desk paths are `/desk/atlas`, `/desk/scout`, `/desk/quill`. After magic-link sign-in you land on the desk. Click the pal. Talk.

Human routing is a feature. It keeps the snack jar honest. If you later add tools, you can still keep the human as the switch — Bix pals do not pretend to control a real desktop.

How credits bite

Bix will not start a reply it cannot afford. The minimum bite is real; the jar is a wallet, not a progress bar that wraps around. If you are modeling cost in tokens instead of snacks, convert with the tokens → $ → packs explainer and the live calculator.

A practical weekend budget:

• 20 thoughtful turns on Saturday ≈ a noticeable dent in 150 starter snacks, not a $200 invoice.
• 80 sloppy regenerations ≈ you will meet the hard stop and learn something.

Design the product so the second path is embarrassing, not default.

Shared jar, separate threads

Threads keep titles. The jar is still shared. That is how a multi-agent desk stays one budget. You can brainstorm with Atlas, research with Scout, and draft with Quill without three wallets. You cannot accidentally open a fourth “free” agent that bills a different cloud.

If you need isolation (client A vs client B), sign in as the right human or wait for stricter workspace features. Do not fake isolation by running three API keys in a notebook and calling it a desk.

What not to automate yet

Do not auto-chain Atlas → Scout → Quill on every message. That triples the bite for questions like “what time is the demo.” Chain only when the artifact changes hands: plan approved, research closed, draft requested.

Do not scrape the live web through a pal that cannot actually browse. Scout will say when it does not have live access. Believe it. Invented citations are more expensive than a snack — they cost trust.

Do not hide the empty-jar state. The nap is the brand. Unlimited products apologize. Bix sits down. Say that on the landing page. We already do, next to pricing.

A sample afternoon that does not melt the key

1:00 Atlas: “Break this marketplace idea into a Friday launch, ten steps, no fluff.” 1:12 You delete steps 7–10 because they were fantasy. 1:15 Scout: “What do we actually know about fees? List assumptions.” 1:25 You write the two facts you do know. 1:30 Quill: “Draft the homepage hero in Bix’s voice. One version.” 1:40 You ship the hero. You do not ask for a ‘more innovative’ rewrite.

You used three pals and maybe a few dozen snacks. An automated graph would have spent that before you finished coffee.

Instrumentation you can do this week

• Watch the jar on every thread. If you cannot see it, you will ignore it.
• Write down snacks per shipped artifact, not snacks per hour.
• Run the calculator with your real messages/day before you buy Pro.
• Keep Stripe packs as the only refill. Side-loading tokens trains the wrong muscle.

When to buy which pack

Starter ($10 / 1,000) is a second weekend. Plus ($25 / 2,750) is a loud launch week. Pro ($50 / 6,000) is when the desk is part of how you write, not a toy. Numbers live on usebix.com/#pricing. The calculator will tell you if your token habit is secretly a Pro habit.

Try the desk like a skeptic

Sign in. Send Atlas a messy goal. Send Scout the same goal and compare the shape of the answer. Send Quill only the final brief. If the three replies feel different, the crew is working. If they feel identical, your prompt is a blob — fix the prompt, not the models.

Then stop. Close the tab. If you still want the pals tomorrow, feed the jar. That is a multi-agent desk that does not burn the budget: people as the router, prepaid as the fuse, three named jobs instead of an agent swarm.

A budget script you can run every Monday

Write four lines in a note, not a dashboard:

• Sends last week.
• Packs bought last week.
• Artifacts shipped (heroes, emails, FAQs, plans).
• One prompt you will never send again.

If sends climbed and artifacts did not, you used the desk as a fidget toy. Atlas can help you name a smaller job. If packs climbed and artifacts climbed, you are fine — buy Plus or Pro on purpose from pricing. If you cannot remember the numbers, you are back to an invisible API key. The calculator and the tokens essay exist so Monday stays boring.

Invite a teammate only after you can say those four lines out loud. A second human on a confused jar doubles the fidget. A second human on a named crew doubles the shipping. Sign in, run one week solo, then decide. Multi-agent is a staffing choice, not a framework checkbox.

The one-sentence architecture

A prepaid multi-agent desk is a visible wallet, three named jobs, a human switch, and a hard stop. Everything else — graphs, orchestrators, “autonomous” loops — is optional and usually a snack tax. If you cannot draw that sentence on a napkin, you are not ready for a fourth pal. Draw it, meet the crew, and keep the calculator next to the launch checklist. The budget you save is the product. When someone asks for a demo video of “agents talking to agents,” show them a screen recording of you clicking Atlas, then Scout, then Quill. It is less cinematic. It is also how a weekend company keeps the lights on. Sign in and record that once for your own team. If the recording is longer than three minutes, you added a swarm. Cut until a stranger can name the three jobs. Then stop talking about agents and go ship the thing the pals were supposed to help you write.