Agents and Skills

Bundled agents

trusty-agents ships with these agents in .trusty-agents/agents/:

Run tagent agents list for the live list.

Writing a custom agent

1. Drop a TOML file into .trusty-agents/agents/<your-name>.toml:

[agent]
name = "rust-engineer"
role = "engineer"
model = "anthropic/claude-sonnet-4-6"
description = "Rust 2024-edition specialist"

[llm]
temperature = 0.2
max_tokens = 8192

[tools]
allowed = ["read_file", "write_file", "list_dir", "grep_files", "shell_exec"]

[skills]
include = ["rust-async-tokio", "rust-error-handling"]

[system_prompt]
content = """
You are a senior Rust engineer specializing in async Tokio applications…
"""

2. (Optional) Use a starter from .trusty-agents/agent-templates/:

cp .trusty-agents/agent-templates/engineer.toml .trusty-agents/agents/rust-engineer.toml
$EDITOR .trusty-agents/agents/rust-engineer.toml

3. Verify it loads:

tagent agents list | grep rust-engineer

4. Invoke it:

tagent --direct rust-engineer --task "Write a tokio TCP echo server"

Bundled skills

Skills are Markdown files with optional YAML frontmatter. By default, discovery scans two local sources (in priority order):

1. <project>/.trusty-agents/skills/
2. ~/.trusty-agents/skills/

Add more sources — extra local directories or remote git repositories — via .trusty-agents/skill-sources.toml; see configuration.md.

tagent skills list
tagent skills sources    # which directories were scanned

Writing a custom skill

A skill is a single Markdown file:

---
name: rust-async-tokio
description: Tokio async patterns for Rust 2024
tags: [rust, async, tokio]
---

# Tokio Async Patterns

## Structured concurrency

Use `tokio::spawn` and join handles instead of bare `tokio::spawn` + drop:

…

Drop it in any of the discovery directories. Reference it from an agent's TOML:

[skills]
include = ["rust-async-tokio"]

Or load it dynamically inside an agent turn via the skill_loader tool.

How injection works

The SystemPromptBuilder layers content in this order before sending to the LLM:

[base system prompt from agent TOML]

[CLAUDE.md walked from CWD up to project root]

[# Skill: <name>\n<content>] for each resolved include

Skills resolved at runtime by skill_loader(skill_name) are appended to the running conversation as tool-result messages, so they don't bloat the system prompt for unrelated turns.

Tools available to agents

The default tool registry exposes these to most engineering agents (see the [tools].allowed list in each agent TOML for per-agent restrictions):

• File I/O: read_file, write_file, list_dir, grep_files
• Shell: shell_exec (allowlisted; gated to local-ops-agent and qa-agent)
• Web: web_search (Brave), fetch_url
• Memory: memory_search, vector_search, kuzu_recall, store_memory, retrieve_memory, list_memory_keys
• Skills: skill_loader, skill_list
• Code search: search_code, search_memory, search_skills
• Workflow: phase_audit, finish_task
• Tickets: create_ticket, get_ticket, close_ticket
• Format conversion: format_translator (Markdown ↔ HTML, JSON ↔ TOML, YAML ↔ JSON)
• Delegation (PM only): delegate_to_agent
• CTRL: start_pm, list_projects, task_status, self_project_status, initiate_self_task, search_docs, memory_store, memory_recall

Discovering everything from the CLI

tagent agents list
tagent skills list
tagent skills sources

Or, from inside CTRL, just ask:

CTRL> what agents are available?
CTRL> how do skills get injected into prompts?

The search_docs tool will surface the relevant docs.