Skip to main content

Design agent templates and souls

An agent template defines a reusable responsibility: the kind of agent, its inference profile, operating directives, report format, and optional personality. A soul holds the reusable voice, tone boundaries, coaching style, and anti-sycophancy policy. Keeping those concepts separate lets one personality improve consistently across several templates without copying instructions between them.

Review the template catalog​

Open Settings → Agents → Agent Templates. The catalog shows the template kind, model identity, active version, and whether a review needs attention.

All screenshots
Each Project Waddle template has a narrow responsibility and visible version instead of one all-purpose agent with an unreviewable prompt.
ScreenshotMobile view
Lotti agent templates for orbital habitat safety, Project Waddle day planning, sardine supply monitoring, and fish diplomacy coaching

The built-in kinds serve different scopes:

  • Task Agent works against one task and its linked evidence.
  • Day Agent plans and revises a day.
  • Project Agent follows a project over time.
  • Event Agent follows an event workspace.
  • Template Improver supports the versioned one-on-one review process.

Edit a template​

Select a template to edit its current configuration. Choose a default inference profile with the tools and modalities the responsibility requires, then optionally assign a soul.

All screenshots
The sentinel's operational directive describes what to inspect; Admiral Pebble contributes the communication and challenge style.
ScreenshotMobile view
Orbital Habitat Sentinel template editor using Project Waddle Command and Admiral Pebble with habitat safety and report directives

Use the two directive fields deliberately:

  • General Directive defines the responsibility, boundaries, and decision rules.
  • Report Directive defines what a useful saved report must contain.

Saving an edited template creates a new version, and existing history remains available for audit and rollback. Avoid hiding unrelated responsibilities in one long directive just to reduce the number of templates.

Run a template one-on-one​

The template 1-on-1 page frames proposed changes against real wake activity, feedback, performance, and previous sessions. A pending session is a proposal, not an automatic rewrite.

All screenshots
The example proposal responds to a concrete failure: the safety finding was correct, but the go/no-go decision was buried under fish trivia.
ScreenshotMobile view
Orbital Habitat Sentinel one-on-one showing a proposal about surfacing the launch decision and recent performance metrics

Start the conversation, inspect the evidence, and approve only the changes that improve the responsibility. The approved result becomes another versioned template record.

Keep personality reusable with souls​

Open Souls to see the personalities that templates can share. A soul is not another task prompt; it controls how an agent communicates, coaches, and pushes back.

All screenshots
Project Waddle separates a calm flight director, a curious scientist, and a practical cargo coordinator from the jobs that use them.
ScreenshotMobile view
Lotti soul catalog containing Admiral Pebble, Dr. Flipper, and Captain Sardina

The soul editor keeps four concerns explicit:

All screenshots
Admiral Pebble can stay warm and decisive while still challenging optimistic launch assumptions with evidence.
ScreenshotMobile view
Admiral Pebble soul editor with voice, tone bounds, coaching style, and anti-sycophancy policy
  • Voice Directive describes the recognizable voice and communication priorities.
  • Tone Bounds state what the voice must avoid.
  • Coaching Style describes how it helps someone reach a decision.
  • Anti-Sycophancy Policy explains when and how it should disagree.

Use concrete behavioral boundaries. “Be helpful” is hard to review; “challenge optimistic launch assumptions and cite the observation that changed your conclusion” can be checked against actual output.

Evolve a shared soul carefully​

A soul 1-on-1 draws feedback from every template using that personality. That makes the change powerful: approving it can improve several agents, but a poor change can also affect all of them.

All screenshots
The sharing count and session history make the blast radius visible before personality changes are approved.
ScreenshotMobile view
Admiral Pebble soul one-on-one with a proposal, one sharing template, and completed session history

Read the sharing count, proposed feedback summary, and previous recap before starting a personality review. If feedback applies only to one responsibility, change that template instead of teaching the shared soul a task-specific rule.

Return to Run and monitor agents to inspect the instances and wake cycles produced from these blueprints.