Skip to main content

​
Agents

An agent is an AI entity with a role, a model, and tools. It persists memory across sessions and executes tasks autonomously.
  • name — unique identifier
  • role — what the agent does
  • model — LLM provider and model (e.g. anthropic:claude-sonnet-4-5)
  • tools — what the agent can do (read, write, bash, http_fetch, etc.)
  • memory — persistent context across sessions (shared + per-agent)
  • identity — persona, responsibilities, communication style
  • skills — installable knowledge packs that teach new domains

​
Teams

Agents are organized into teams. Teams are just grouping — they don’t restrict what agents can do. A project can have multiple teams with different specializations.

​
Tasks

A task is a unit of work assigned to an agent. The agent executes it in an isolated sandbox with access to its configured tools. Tasks have a lifecycle: pending → assigned → in_progress → review → done or failed. When a task fails, Polpo can automatically retry, fix, or escalate based on your retry policy.

​
Missions

A mission is a multi-step workflow composed of tasks. Missions define:
  • Tasks with dependencies (task B waits for task A)
  • Checkpoints — pause for human review before continuing
  • Quality gates — automated pass/fail based on assessment scores
  • Delays — timed waits between task groups
  • Volatile agents — temporary specialists created for the mission

​
Assessment

Every task result is automatically assessed:
  1. Agent completes the task
  2. Assessment runs expectations (tests, file checks, LLM review)
  3. G-Eval scoring across weighted dimensions
  4. Pass → done. Fail → fix prompt → retry

​
Memory

Two levels:
  • Shared memory — accessible by all agents in the project
  • Agent memory — private to a single agent, persists across sessions
Memory is injected into the agent’s system prompt before every response.

​
Tools

Agents interact with the real world through tools: read/write files, run commands, fetch URLs, send emails, browse the web, analyze images, and more. Tools run inside the agent’s sandbox with access to the project volume.

​
Volumes

A volume is a FUSE-based persistent filesystem for a project. Every project gets one automatically. All agents share the same volume — when one agent writes a file, the next agent sees it. Memory, skills, and shared context live on the volume.

​
Apps

An app is the main unit of infrastructure for coding agents. It bundles a codebase, servers, and ports into a managed environment. One call gives you a live preview, IDE, and terminal. Apps can hibernate (stop compute, keep files) and wake instantly.

​
Codebases

A codebase is the code layer inside an app — project source, git history, dependencies, and builds. Unlike volumes (FUSE-based, shared), codebases use local block storage suitable for running dev servers and builds.

​
Servers

A server is a long-running process inside an app — dev servers, terminals, IDEs. Servers stay alive between agent sessions and restart automatically after hibernation.

​
Templates

A template is a pre-built snapshot that apps are created from. It includes a runtime, project code, installed dependencies, and default server definitions. Apps boot from templates in seconds.

​
Chat Completions

Every agent is accessible via an OpenAI-compatible API. Send messages, get streaming responses. Pass an appId to run the agent inside a persistent app. Works with any OpenAI client library.

​
Vault

Encrypted credential storage for agents. API keys, SMTP credentials, OAuth tokens — stored with AES-256-GCM encryption, decrypted only at runtime inside the sandbox.