Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Libraryworkflows / 2026-06-22

I run 100-agent workflows on a budget model. Here's the catch.

One reported 235-agent run used a budget model inside Claude Code's workflow harness. Keep provider compatibility, verification and the cost of lower-quality answers in the comparison.

6 parts / Hover to explore. Click to read.

One reported 235-agent run used a budget model inside Claude Code's workflow harness. Keep provider compatibility, verification and the cost of lower-quality answers in the comparison.

Keep the coordination + checks
               |
      choose compatible workers
               |
       bounded comparison
               |
     findings + cost + misses
               |
       human review of trade-off
A cheaper worker changes one part of the surrounding workflow. Compare its findings and verification burden as well as token usage; shared agreement and lower prices do not establish correctness.
A System 7 concert scene: one conductor labelled CLAUDE CODE, an orchestra of identical players badged GLM and MINIMAX.
Keep the coordination. Choose the model for the work it has to do.

In the run behind this June 2026 essay, I dispatched 235 agents to review a codebase. Claude Code provided the harness. A budget model filled the worker seats.

That distinction is useful when you look at the bill. The model supplies reasoning and responses. The harness supplies the surrounding execution: tools, dispatch, coordination and the path by which results come back.

Know which part you are changing

Claude Code's dynamic workflows generate orchestration scripts for a task. Its ultracode setting combines xhigh effort with automatic decisions about when to use a workflow. Anthropic documents both the feature and its higher usage in the dynamic workflows introduction.

For this review, the useful shape was to split the work, challenge the findings with other agents, then assemble the surviving results. Those stages were part of the workflow around the model. They weren't a special ability available only inside one set of model weights.

Automatic fan-out also makes it easy to spend a lot of tokens. A cheaper call repeated hundreds of times is still hundreds of calls.

Keep the conductor and test another orchestra

The conductor stays fixed. Swap the orchestra, and the bill collapses.

I used the same coordination structure with a cheaper provider underneath it. GLM and MiniMax were the alternatives discussed in this setup. Claude was one possible model for a seat rather than the definition of the whole workflow.

Provider substitution still has to work with the harness and its tools. Treat it as a configuration to test, not a promise that every provider supports every workflow feature unchanged.

Stacked diagram: HARNESS (workflows, dispatch, tools) above a swappable MODEL slot branching to Claude, GLM, MiniMax.
The harness owns the coordination. A compatible model fills the seat.

Keep the quality trade visible

In this setup, a cheaper worker gave me less depth per agent. I wasn't treating one GLM answer as interchangeable with one Opus answer. I was buying more independent attempts within the budget and spending some of that budget on verification.

The original “fifty for the price of one” comparison described that breadth argument. It wasn't the measured ratio for the run below. Parallel reviewers can catch each other's mistakes, but agreement between them doesn't establish that they're right. They can share a blind spot.

I still reviewed the results. A lower token price made exploration more affordable; it didn't remove the cost of checking it.

Read the run's numbers at their proper scale

The reported run used:

  1. 235 agents across the workflow.
  2. 13.12 million tokens between them.
  3. An estimated cost of roughly one tenth of the premium-rate comparison.
  4. The same fan-out, verification and synthesis structure.
provider cost dial

The orchestra (model)

Agents in the fan-out: 235

Run cost: $117.50

Real published rates, Jun 2026 ($/1M in/out): Claude (Opus 4.8) $5/$25, GLM-4.6 $0.43/$1.74, MiniMax M2 $0.26/$1.00. Assumes ~56k tokens per agent (45k in, 11k out).

Those are figures from that run, not a current price quote or a general tenfold saving. A different model, task, retry rate or verification burden can change the result.

Try one bounded comparison

Start with a task whose output you can judge. Use a dynamic workflow where parallel work is useful, configure a compatible budget provider, and keep the acceptance checks the same.

Inspect the findings as well as the token usage. Did the cheaper workers miss important evidence? Did the review stage catch it? How much human correction remained?

Then decide which seats the model has earned. Keep the coordination you understand and change one part at a time. The useful outcome is good work at a cost you can sustain, not a large agent count to admire.

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 8