Control agent costs
Project lead · Build the agent team
Steps at a glance
1. Know what you are actually paying for
Buzz itself is free. The cost is entirely in the models: either a harness subscription or per-token API usage. What changes with Buzz is the rate at which you consume it — a team of agents working in parallel across several channels burns through budget far faster than one person in one chat window.
Real-world experience puts the framing bluntly: today's $200 a month can become $2,000, and subsidised subscription plans are widely expected to get more expensive.
2. Pin cheap models to cheap work
The single most effective lever. Give your reasoning-heavy agents a strong model and pin everything else — summarising, formatting, routing, checking — to a cheap one.
In practice this means deciding per agent rather than per workspace, at creation time. See Create and configure an agent.
The mental model: token budgeting, not token maxing. A cheaper model doing a simple task well is not a compromise.
3. Cap parallelism
Agent settings include a parallelism control — the number of concurrent sessions an agent may run. Higher values get more work done at once and multiply spend at the same rate.
Warning
Whether this setting is fully honoured has been questioned. Verify it with a small test before you rely on it for budgeting — that makes it a cost question, not a curiosity.
4. Use shared compute where it fits
If your community shares compute, agents can run against a local model with no API key and no per-token cost. That suits the high-volume, low-stakes half of an agent team well.
Understand the trade before adopting it: prompts run on community members' hardware. See Share and use shared compute and Buzz Mesh and where prompts go.
5. Watch the spend
Per-turn token and cost accounting exists at protocol level, reported privately to the agent's owner. Whether any of it surfaces in the app varies by build — if you cannot find it, track spend at your provider's usage dashboard.
Set yourself a review point — weekly during an active project — and a rough ceiling you would want to know about crossing. Without one, the first signal is the invoice.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026