How Your Agent Finds the Right Tools, Agents and Skills
How agents use tool search and shortlists to find the right tools, agents and skills without carrying the whole catalogue.
Overview
Modern AI models can "think" before they answer, and the harder they think, the longer you wait. Thinking hard helps on a tricky analysis, but it makes a simple question like "what's the capital of France?" slower for no benefit.
The effort preprocessor takes care of this for you. It is on for all agents. Before each message reaches your agent, Sapience asks Jev (a fast decision model from TypeSafe) how much reasoning effort the message deserves: low, medium or high. Simple questions get quick answers. Hard ones still get the full treatment.
There is nothing to switch on or configure.
What You See
You see nothing extra on screen. What changes is the wait: a simple question comes back sooner, and a hard one takes as long as it needs.
What Happens Behind the Scenes
- You send a message to an agent.
- Before the agent starts, Jev reads your message (and the agent's last reply, so a short follow-up like "yes, do it" is judged in context) and picks low, medium or high. This takes a fraction of a second.
- The agent answers using that level of reasoning effort.
- On your next message, Jev decides again, so each message gets the effort it needs.
Your message | Typical pick |
A quick factual question or a short chat | low |
A normal drafting or summarising request | medium |
A long, multi-step or analytical request | high |
When Jev Is Not Asked
Your own choices always win:
Situation | What happens |
You (or the request) already set a reasoning effort | Your setting is used. Jev is not asked. |
A skill pins its own effort | The skill's effort is used. Jev is not asked. |
Jev is temporarily unavailable | The agent's own configured effort is used, and your message is answered normally. |
If your agent hands work to another agent, that agent gets its own pick, so a simple delegated task is not slowed down by the main agent's effort.
What You Can Expect
We measured this on the same 13 real prompts, comparing an agent whose usual effort is high on every message with the effort preprocessor's picks:
Usual effort | With the effort preprocessor | |
Median time to the full answer | 27.4 s | 25.1 s |
Median time to the first words | 23.9 s | 18.8 s |
Simple prompts, time to full answer | 24.2 s | 11.4 s |
Answer quality (judge score out of 5) | 3.6 | 3.9 |
Good to Know
- Long, heavy requests are not faster. They are given high effort, as before.
- It is automatic. There is no screen setting for it. If you need a particular effort for a message, set it yourself and Sapience will respect it.
- Jev's decision is quick. It adds a fraction of a second, far less than it saves on simple questions.
- It adds no meaningful cost. Jev's decision costs about a hundredth of a cent per message, and quick answers use less reasoning.
Summary
The effort preprocessor is on for all agents. It lets Jev choose low, medium or high reasoning effort for each message, so simple questions come back faster while hard questions keep their full effort. Your own settings and a skill's pinned effort always take priority, and if Jev is unavailable your agent simply uses its usual effort.