If you run OpenClaw with multiple agents, you've probably hit the same frustration: one config to rule them all. A coding agent that needs deep workspace context and a lightweight triage bot that should stay lean have been forced to share the same agents.defaults bootstrap settings — until now.
A new PR merged today (#82036, by @BunsDev) introduces per-agent bootstrap profiles, letting you override context injection behavior and bootstrap size limits at the individual agent level.
What Changed
Three new optional fields are now available inside each agents.list[] entry in your OpenClaw config:
contextInjection— controls which context files are injected at bootstrap for this specific agentbootstrapMaxChars— caps the character budget for a single bootstrapped file for this agentbootstrapTotalMaxChars— caps the total bootstrap context size across all files for this agent
When any of these is set on an agent entry, it takes precedence over the corresponding value in agents.defaults. If unset, the defaults apply as before — fully backward compatible.
Why It Matters
Multi-agent setups often have wildly different needs. Consider a typical homelab setup with two agents:
- A coding agent that needs your full AGENTS.md, SOUL.md, USER.md, recent memory files, and a generous bootstrap budget to do useful work
- A triage/heartbeat agent that should spin up fast with minimal context, process a lightweight check, and get out
Before today, both agents shared the same bootstrap profile. You could either over-provision the lightweight agent (wasting tokens on every heartbeat run) or under-provision the coding agent (losing important workspace context). Now you can tune each independently.
How to Configure It
Here's an example openclaw.json showing per-agent overrides alongside your defaults:
{
"agents": {
"defaults": {
"bootstrapMaxChars": 40000,
"bootstrapTotalMaxChars": 120000
},
"list": [
{
"id": "main",
"contextInjection": {
"files": ["SOUL.md", "USER.md", "AGENTS.md", "MEMORY.md"]
},
"bootstrapMaxChars": 60000,
"bootstrapTotalMaxChars": 200000
},
{
"id": "triage",
"contextInjection": {
"files": []
},
"bootstrapMaxChars": 8000,
"bootstrapTotalMaxChars": 20000
}
]
}
}
The triage agent above starts with zero injected files and tight size limits — perfect for a cron-driven heartbeat that doesn't need your full workspace identity loaded.
Improved Diagnostics
The PR also threads the resolved session agent ID through the /context diagnostic output, so when you run /context in a session, the report now accurately reflects the bootstrap profile actually applied to that agent — not just the global defaults. This makes it much easier to verify your per-agent settings are working as expected.
Under the Hood
The implementation adds the three fields to AgentEntrySchema in src/config/zod-schema.agent-runtime.ts, updates the central bootstrap file resolver and bootstrap limit resolver to check per-agent overrides before falling back to defaults, and passes the resolved agent ID through embedded, compact, CLI, and diagnostic code paths. Focused regression tests cover the resolver logic and the /context report for session-agent bootstrap limits.
No existing behavior changes if you don't add these fields — your current config works as-is.
What to Expect Next
This lands in a future v2026.5.x stable build. If you're tracking the beta channel, it should appear in the next beta tag cut after the PR clears CI and maintainer review. Keep an eye on the GitHub releases page for the tag that includes PR #82036.
Per-agent bootstrap profiles are a quiet but meaningful unlock for anyone running more than one agent — and given how common split coding/utility setups have become, this one's been a long time coming.
Source: PR #82036 — support per-agent bootstrap profiles by @BunsDev