Smart Context Management: Start Small

Everything in an agent’s instructions and tool list is paid for on every request, whether the task needs it or not. This group of Smart Context Management is about not loading it in the first place.

Skills load on demand

The model sees each skill’s name and description. The full instructions load only when it calls load_skill, and stay until unload_skill. An agent with many skills attached pays for the one or two a task actually uses. I covered setting skills up in an earlier post: Skills: Keep the Agent Prompt Small, Load the Playbook When Needed.

The docs add a tip I’d underline: keep SKILL.md as an orientation file and put the bulk in references/, which the agent reads only when a topic comes up.

The tool list is filtered

Tool definitions are part of every request. Solace Agent Mesh leaves out tools the model can’t use yet, such as tools owned by a skill that hasn’t loaded. With role-based access control configured, it also leaves out tools and peer agents the user isn’t allowed to call. Fewer schemas, smaller requests.

Peer agents work in their own context

When an agent delegates to a peer, it doesn’t get the peer’s conversation or tool results back. It gets:

  • the peer’s final reply
  • the names, versions and status of artifacts the peer produced
  • task status and metadata, such as token usage

If it needs an artifact’s content, it calls load_artifact. The delegating agent sees the conclusion, not the working.

How I’d apply it

  • Move long playbooks from the system prompt into skills.
  • Give each agent only the tools its role needs, and delegate the rest to a peer.
  • Write skill descriptions around the problem, since that one line decides whether the skill loads.

None of these show up on the Smart Context Savings card, because Agent Mesh can’t measure what the model would have received without them. They still cut every request.

Docs:

Blog post:

Download the Solace Agent Mesh desktop app and check how many tools your busiest agent carries on every call.

Which would shrink your agents more: splitting the prompt into skills, or splitting the agent into peers?

1 Like