Base Agent Templates¶

Pre-built .agent files in examples/bases/ that you extend instead of writing from scratch.

Available bases¶

examples/bases/
  claude.agent           ← Claude Code (general-purpose)
  claude-noether.agent   ← Claude + Noether composition skills
  gemini.agent           ← Gemini CLI (general-purpose)
  gemini-noether.agent   ← Gemini + Noether
  codex.agent            ← Codex CLI (OpenAI)
  codex-noether.agent    ← Codex + Noether
  local.agent            ← Local Ollama (no API keys)
  local-noether.agent    ← Local + Noether

Extend a base¶

# my-pricing-analyst.agent
apiVersion: agent/v1
name: ota-pricing-analyst
version: 1.0.0
description: "Analyzes hotel rate data, detects anomalies"

base: ./bases/claude-noether.agent

merge:
  skills: append
  tools: append
  behavior: append
  trust: restrict

# Add domain-specific skills on top
skills:
  - data-analysis

tools:
  mcp:
    - postgres
        connection: env.OTA_DB_URL
        mode: read-only

behavior:
  traits:
    - cite-everything    # for data sources
  temperature: 0.1       # stricter than parent's 0.2

# Tighter sandbox than parent
trust:
  scope: [./src, ./tests, ./reports]

Why bases matter¶

Consistency — every agent in a team uses the same baseline behavior, trust, and runtime preferences
Less duplication — define common config once
Auditability — security teams approve the base, individual agents can only restrict further
Evolution — improving a base improves all child agents

What each base provides¶

`claude.agent` / `gemini.agent` / `codex.agent`¶

reasoning-high capability tier
senior-developer persona
Traits: think-step-by-step, be-concise, self-review
Trust: filesystem: scoped to ./src, ./tests, ./docs
network: allowed, exec: sandboxed
Tools: bash, browser, mcp:github

`local.agent`¶

reasoning-mid (smaller models)
No network access (network: none)
No GitHub MCP (no API keys)
Same trust scopes

`*-noether.agent` (all four)¶

Inherit from their plain base + add:

Skills: noether-compose, noether-run, noether-search, noether-serve
Traits: noether-first, noether-verify
Trust scope: + ./compositions, ./stages

Cookbook¶

A pricing analyst¶

base: ./bases/claude-noether.agent
skills:
  - data-analysis
tools:
  mcp:
    - postgres:
        connection: env.PRICING_DB_URL
behavior:
  persona: pricing-analyst
  temperature: 0.0    # deterministic

A frontend developer¶

base: ./bases/codex.agent
skills:
  - code-execution
tools:
  native:
    - npm
    - playwright
behavior:
  persona: frontend-engineer
  traits:
    - test-first
trust:
  scope: [./src, ./tests, ./public]

A local-only research assistant¶

base: ./bases/local.agent
model:
  preferred:
    - local/llama3:70b
behavior:
  persona: researcher
  traits:
    - cite-everything
    - never-guess

Publishing your own bases¶

Push your bases to the registry so your team can extend them:

agentspec push my-org-base.agent --registry https://registry.mycompany.com

Then in any agent:

base: ag1:abc123def456    # registry hash

Or by name:

base: my-org/standard-base@1.0.0