polpo-agents（波尔波特工）

Polpo Agent Design

Agent Configuration

An agent is defined by its AgentConfig. Create via API or dashboard.

{
  name: "coder",                              // unique identifier
  role: "Senior Full-Stack Engineer",         // human-readable role description
  model: "xai/grok-4-fast",                   // provider/model format
  allowedTools: ["bash", "read", "write", "edit", "glob", "grep"],
  systemPrompt: "You are a senior engineer. Write clean, tested code.",
  maxTurns: 150,                              // max LLM turns per session
  maxConcurrency: 3,                          // max parallel tasks
  reasoning: "medium",                        // thinking depth
}

Model Selection

Format: provider/model. Choose based on task complexity and cost.

Browse all available models at polpo.sh/api/gateway/models (JSON) or see docs.polpo.sh/developers/reference/providers for a guide.

Tools

Tools define what an agent can do. Assign only what's needed — follow the principle of least privilege. Every tool must be explicitly listed in allowedTools to be available.

Coding Tools

• bash — Execute shell commands (sandboxed)
• read — Read files
• write — Create/overwrite files
• edit — Search-and-replace in files
• glob — Find files by pattern
• grep — Search file contents

Additional Tools

Polpo supports additional tool categories (integration, email, image generation, and more). Each must be explicitly assigned in allowedTools — agents have no tools by default beyond what you configure.

For the complete tool catalog, see docs.polpo.sh/docs/agents/tools.

Security Notes

• All tool execution runs inside an isolated sandbox — agents cannot access the host system or other projects.
• vault_get only returns credentials explicitly assigned to that agent. Agents cannot access other agents' credentials.
• email_* tools enforce emailAllowedDomains — restrict which domains an agent can email.
• Follow the principle of least privilege — only assign tools an agent actually needs for its role.

System Prompt

The system prompt defines the agent's behavior. Keep it focused.

You are {name}, a {role}.

{Core instructions — what to do, how to do it}

{Constraints — what NOT to do}

{Output format preferences}

Good system prompt:

You are a code reviewer. Review code for bugs, security issues, and performance problems.
Be concise. Flag critical issues first. Suggest fixes with code examples.
Never approve code with SQL injection or XSS vulnerabilities.

Bad system prompt: Overly long, generic instructions that Claude already knows.

Identity

For agents that interact with humans (email, chat, social), define an identity:

identity: {
  displayName: "Alex",
  title: "Engineering Lead",
  bio: "I help teams ship reliable software.",
  tone: "Direct, technical, no fluff",
  personality: "Pragmatic problem-solver",
  responsibilities: [
    { area: "Code review", description: "Review all PRs", priority: "critical" },
    { area: "Architecture", description: "System design decisions", priority: "high" },
  ],
  socials: { github: "alex-eng", twitter: "@alex_ships" },
}

For more on agent identity, see docs.polpo.sh/docs/agents/definition.

Memory

Memory persists across sessions. Two levels:

Project Memory

Shared by all agents. Use for project context, conventions, architecture decisions.

# Project Context
- Next.js 15 + TypeScript + Tailwind
- PostgreSQL via Drizzle ORM
- Deploy on Vercel

# Conventions
- Use Vitest for tests
- Prefer server components
- No default exports (except pages)

Agent Memory

Private to one agent. Auto-accumulated as the agent works. Use for agent-specific learnings.

# Learned Preferences
- User prefers functional style over classes
- Always run tests before committing
- Use pnpm, not npm

Memory is read by the agent at the start of each session. Update via API:

• PUT /v1/memory — project memory
• PUT /v1/memory/agent/{name} — agent memory

For more on memory, see docs.polpo.sh/docs/agents/memory.

Vault

Store service credentials that agents access at runtime. Credentials are scoped per-agent — an agent can only access entries explicitly assigned to it.

// Store credentials
POST /v1/vault/entries
{
  agent: "emailer",
  service: "gmail",
  type: "smtp",
  credentials: { host: "smtp.gmail.com", port: 587, user: "...", pass: "..." }
}

// Agent uses vault_get("gmail") at runtime — credentials never in system prompt

Credential types: smtp, imap, oauth, api_key, login, custom.

For more on vault, see docs.polpo.sh/docs/agents/vault.

Agent Patterns

See references/patterns.md for multi-agent architectures, specialization patterns, and reporting hierarchies.

Need Help Designing Your Agent?

Not sure which tools, model, or configuration your agent needs? Ask your coding agent:

"I want to create a Polpo agent that [describe what it should do]. Help me choose the right model, tools, and system prompt. Use the polpo-agents skill for reference."

Your coding agent will help you pick the right configuration based on your use case.