Platform Documentation
The foundational concepts behind CadreGPT, explained plainly. Every execution in the workspace operates on these deterministic principles.
Work
Work is a defined business objective with context and success criteria. You define what "done" means before execution starts, so the agent has an unambiguous outcome target.
Task
A task is a discrete unit of work assigned to an agent or DAG graph. Each task carries its own credit budget with a hard stop — execution pauses automatically at the limit.
Run
A run is an active execution turn of a task. Telemetry streams live thoughts, tool calls, and LLM completions directly into the dashboard over Server-Sent Events (SSE).
Agent & Team
Agents are specialist personas with instructions, system prompts, and assigned tool sets. Lead agents coordinate complex multi-step work and delegate to specialist teammates.
Human Approvals
Sensitive tool operations are gated by policy: execution safely pauses into AWAITING_APPROVAL and requests human sign-off before proceeding. Agents cannot bypass approval gates.
Execution Limits
Every task runs under a wall-clock deadline, a per-turn tool-call cap, a per-turn tool-spend ceiling, and a per-run cost ceiling, and is stopped when it crosses one. Limits are enforced by the runtime, not requested in the prompt.
Evidence Verification
Every finished task preserves an immutable audit trail and artifact record in Cloudflare R2, enabling thorough inspection of actual results against expectations.
Evaluation
An evaluation runs your agent against a saved dataset of prompts and expected behaviors, then scores the results so you can compare versions before promoting one to production.
Dry Run
A dry run executes an agent against a sample input in the playground so you can inspect its plan and output before committing the change.
Trace
A trace is the span-level record of a run: every model call, tool invocation, token count, and latency, streamed live and stored so you can review what actually happened.
Knowledge Base
A workspace-scoped store of your documents. Agents retrieve the relevant passages to ground their answers in your own data instead of guessing.
Memory
Memory lets an agent retain context across runs within your workspace. You set an expiry or delete entries whenever you want, and nothing crosses into another tenant.
Guardrail
A guardrail is a workspace policy that bounds agent behavior: block PII, require valid JSON output, hold to approved topics, and cap cost per run before execution.
Agent Version
Every change to an agent is captured as a revision you can diff, promote, or roll back, so a prompt or model edit that made results worse is reversible.
MCP Server
An MCP server publishes tools that CadreGPT discovers and hands to agents, so you connect a new capability without writing a custom connector.
Embed Token
An embed token publishes an agent as a chat widget or a public endpoint, locked to the origins you allow and rate-limited per token.
Handoff
A handoff moves a conversation from an agent to a person in a shared inbox, where teammates claim it, reply in the thread, and resolve it.
Ready to explore further?
See how these building blocks assemble in our product architecture or learn about governance in security and approvals. To start, check out the ai agent builder overview.
Launch Free Sandbox Workspace →