Agents
An agent combines a role, model, prompt, and tool policy. An agent can also have identity files, skills, and persistent memory when those resources are configured.- name — unique identifier within the project
- role — the agent’s responsibility
- model — provider and model identifier, such as
anthropic/claude-sonnet-4-5 - allowedTools — the tools the agent may invoke
- memory — optional shared and agent-specific persistent context
- identity — persona, responsibilities, and communication guidance
- skills — installable instructions and supporting resources
Teams
Agents belong to teams. Teams group agents for orchestration and task assignment; they are not an authorization boundary.Tasks
A task is a persisted unit of work that can be assigned to an agent and executed with the project’s runtime and tools. The task status values are:draft, pending, awaiting_approval, assigned, in_progress, review, done, and failed.
The exact path through those states depends on assignment, approvals, assessment, and retry policy. A task does not necessarily visit every state.
Missions
A mission is a multi-step workflow that creates and coordinates tasks. A mission can define dependencies, task groups, checkpoints, quality gates, delays, and temporary specialist agents.Agentic Loops
An Agentic Loop is a named graph of steps. Steps can run agents, invoke tools, pause for human input, branch on previous output, and apply permission rules. Loops are stored separately from missions and can be invoked through the API or SDK.Assessment
Assessment runs when a task defines expectations or metrics. Depending on configuration, it can execute tests or file checks, request an LLM review, calculate scores, and trigger a fix or retry path. Tasks without assessment configuration are not automatically scored.Memory
Polpo supports two persistent memory scopes:- Shared memory — project context available to agents
- Agent memory — context associated with one agent
Tools
Tools let agents read and write files, execute commands, make network requests, use connected services, and perform other actions.allowedTools limits which tools an agent can call. Cloud executions acquire a managed sandbox when isolation or compute is required; the open-source runtime uses its configured local execution environment.
Sandbox isolation is a runtime policy. Agents, tasks, and individual chat/task execution requests can set sandbox.isolation to reuse or fresh. reuse favors warm execution; fresh asks the host for a clean sandbox for that run. See Runtime Sandboxes.
Project files
Cloud project files are stored in project-scoped S3-compatible object storage. The Files API, dashboard, and file tools expose that storage through logical roots such asworkspace. Cloud sandbox paths can mount the same project storage for execution.
Volume records currently carry a file-drive or code-drive mode and sync metadata, but both modes use the same Cloud object-storage backend. See Storage.
Chat completions and sessions
Agents are available throughPOST /v1/chat/completions. The response can stream using the OpenAI-compatible chunk format. A request without x-session-id creates a new chat session; send the returned session ID on the next request to continue it.