⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 5 min read

OpenCode Subagents: Choosing Models, Step Caps and Running Benchmarks

How to pair OpenCode subagents with the right models and run fair parallel benchmarks.

Parallel gold lanes on a benchmark grid over dark navy

OpenCode subagents split one coding session into parallel child sessions, and this benchmark-style guide looks at how to pair each subagent with the right model and limits.

Book a free SEO strategy session →

Since v2.3.0 every subagent runs in the background, so the model you give each helper now shapes speed and cost across several jobs at once.

To see agent setups applied to SEO work, book a free SEO strategy session.

Details come from OpenCode's agents documentation and release notes.

What v2.3 changed for benchmarking

The v2.3.0 notes say subagents return immediately, run in the background and report when idle, from PR 63382.

For anyone timing agent workflows, that means wall-clock time is now set by the slowest child, not the sum of all of them.

v2.3.1 moved timed-out shell commands to the background and made a foreground timeout: 0 mean 120 seconds, which affects long test runs inside benchmarks.

Model per subagent

Set model as provider/model#variant; with none set, a subagent inherits the parent session's model.

Subagent jobModel choiceReason
Read-only search (Explore)Inherit, or a fast variantHigh volume, low reasoning
Code reviewA strong reasoning variantMisses are expensive
Changelog or docs draftsA fast variantSimple writing task
Multi-step research (General)InheritNeeds broad context
---
description: Reviews diffs for regressions
mode: subagent
model: anthropic/claude-sonnet-5-5#high
steps: 10
---

Review the current changes for regressions. Report findings
in severity order with file and line references.

Steps as a cost limit

steps sets a maximum number of model steps; on the final step OpenCode removes tools and asks the model to summarise.

In benchmarks, a step cap keeps runaway helpers from skewing results, and new user input resets the allowance.

Built-in baselines

AgentModeUse as baseline for
exploresubagentRead-only search tasks
generalsubagentMulti-step research; cannot launch more subagents
buildprimaryDefault coding agent
planprimaryPlanning without edits

Permissions for fair comparisons

Give every helper in a comparison the same permissions so results reflect the model, not access; OpenCode's permissions system uses ordered rules where the last match wins.

Book a free SEO strategy session →

{
  "agents": {
    "bench-review-a": {
      "mode": "subagent",
      "model": "anthropic/claude-sonnet-5-5#high",
      "permissions": [{ "action": "edit", "resource": "*", "effect": "deny" }]
    },
    "bench-review-b": {
      "mode": "subagent",
      "permissions": [{ "action": "edit", "resource": "*", "effect": "deny" }]
    }
  }
}

Agent B has no model set, so it inherits the parent's model, giving you a clean A/B comparison.

Running a parallel benchmark

  1. Define two or three subagents with identical prompts and permissions but different models.
  2. Ask the parent to launch them all on the same task in one prompt.
  3. Record time to report, steps used and quality of findings.
  4. Wait for final reports; a child that says it is still running has not finished.

Pitfalls

  • V1 fields such as maxSteps are deprecated; see the migration guide.
  • hidden is visibility only, not a restriction.
  • Long shell commands may now continue in the background after a timeout; check the tools docs.
  • Agent definitions merge in configuration order, as the config docs describe, so a later file can override your benchmark settings.

For tested agent setups aimed at business work, see the AI Profit Boardroom.

Book a free SEO strategy session →

Pair each helper with the right model and step cap, and OpenCode subagents become both faster and cheaper; to see agents applied to SEO, book a free SEO strategy session.

FAQ

Do OpenCode subagents run in the background?

Yes. Since v2.3.0 they always run in the background, return their session ID immediately and report to the parent when they go idle.

Which subagents come built in?

General, for research and multi-step work (it cannot launch more subagents), and Explore, which searches and reads without editing files.

What model does an OpenCode subagent use?

The one set in its model field as provider/model#variant, or the parent session's model if none is set.

How do I benchmark two models with subagents?

Define two subagents with identical prompts and permissions but different models, launch both on the same task and compare reports.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week