CloudHarness primitives

Each primitive ships with an API and a SKILL.md. Progression levels go from foundational (P0) to production integration (P6).

P0 — Foundation

Discovery

Machine-readable primitive and skill catalog so agents can discover capabilities and documentation dynamically.

implemented P0

Artifact

Versioned, shareable objects: text, JSON, HTML, files, and blobs.

implemented P0

API Key

Self-service key provisioning; each key is its own account with free credits.

implemented P0

Share

Signed links and email-based sharing for artifacts and (future) runs.

implemented P0

Credits

Per-account credit ledger and usage-based billing for requests, writes, storage, and egress.

implemented P0

Rate Limiting

Per-account fixed-window rate limiting.

implemented P0

Durable Workbox

Durable, account-scoped work items with leased execution, bounded retries, and artifact-backed outcomes.

implemented P0

Queue

Durable FIFO task queue with worker leases, retries, visibility timeout, and dead-letter handling.

selected P0

Run

Observable execution trace for an agent: status, events, lineage, and artifact references.

selected P0

Stack

Durable, account-scoped LIFO collection with atomic push, pop, and peek operations.

selected P0

Trigger

Activation conditions: webhooks, cron schedules, events, and manual fire.

planned P0

Locking / Concurrency

Lease-based distributed locks for shared resources, with ownership tokens and automatic expiry.

selected P0

P1 — Core orchestration

Priority Queue

Durable leased queue ordered by priority and then FIFO sequence within each priority lane.

selected P1

Deque

Durable double-ended collection with atomic push and pop operations at either end.

selected P1

Task Graph / DAG

Graph primitive where nodes are tasks and edges are dependencies, with fan-out, fan-in, and conditional branching.

planned P1

State Machine

Workflow state-machine primitive for long-running, event-driven processes with pause/resume.

planned P1

Scheduler / Cron

Cron-based schedule definitions for tasks, graphs, and triggers.

planned P1

Job Lifecycle API

Unified lifecycle controls: cancel, pause, resume, retry, replay, and inspect across tasks, graphs, and workflows.

planned P1

Structured Output / Schema Validation

Enforce JSON Schema on LLM/tool outputs and retry/repair on validation failures.

planned P1

Checkpoints & Resumability

Save and restore mid-run state for tasks, graphs, and runs.

planned P1

Trace / Lineage Store

Full execution trace of LLM calls, tool calls, events, and state transitions per run.

planned P1

Event Stream / Subscription

Topic-based pub/sub over SSE/WebSocket with filtering, cursors, and replay.

selected P1

P2 — Memory & context

Thread Store

Short-term conversation/thread storage per agent session with pagination, truncation, and rolling summarization.

planned P2

Vector Store

Long-term embedding-based memory with semantic search.

planned P2

Working Memory / Key-Value State

Structured scratchpad for agents with namespacing, TTL, and compare-and-swap support.

selected P2

Knowledge Graph

Entity/relation/observation triple store with extraction helpers and traversal queries.

planned P2

Document Ingestion

Parse, chunk, embed, and index documents (PDF, text, markdown, HTML, JSON).

planned P2

P3 — Tools & actions

Tool Registry

Schema-driven catalog of tools with JSON Schema inputs, versioning, deprecation, and discovery.

planned P3

Tool Executor

Invoke registered tools with schema validation, timeouts, retries, and circuit-breaker policy.

planned P3

Code Execution Sandbox

Safe, ephemeral execution of untrusted code with stdout/stderr capture and resource limits.

planned P3

Browser / Computer-Use Driver

Headless browser actions for navigation, interaction, screenshots, and extraction.

planned P3

MCP / External Tool Bridge

Expose CloudHarness tools via MCP and consume external MCP/OpenAPI tools.

planned P3

P4 — Perception & events

Event Ingestion

Normalize, verify, and deduplicate events from webhooks, queues, polling, and files.

planned P4

File & Media Parsing

Extract content from PDF, DOCX, images (OCR), audio, HTML, JSON, CSV.

planned P4

Observation Log

Immutable append-only log of events, parsed documents, tool outputs, and user messages.

planned P4

P5 — Coordination & multi-agent

Agent Registry

Register and discover agents with capabilities, versions, health, and status.

planned P5

Agent-to-Agent Messaging

Async request/reply and pub/sub messaging between agents with durable inboxes.

planned P5

Delegation / Planning

Hand off goals to other agents and track sub-task completion with result aggregation.

planned P5

Human-in-the-Loop

Approval gates, clarifications, and overrides with notifications and resume API.

planned P5

P6 — Deployment & integration

Webhook / API Gateway

Expose agents/workflows as HTTP endpoints with routing, validation, auth, and rate limits.

planned P6

CLI / SDK

TypeScript and Python clients plus a CLI for authentication, task submission, and deployment.

planned P6

Observability Hooks

Metrics, traces, structured logs, and alerting hooks for the harness and runs.

planned P6