Platform Architecture

Engineered for Autonomous Reliability

Inside the CadreGPT execution runtime: Redis locks, state machine transitions, metered tool dispatch, and tamper-proof evidence trails.

1. Strict State Machine

Tasks follow deterministic states (PENDING → RUNNING → AWAITING_APPROVAL → COMPLETED / FAILED). Invalid transitions are mathematically rejected.

2. BullMQ & Redlock Concurrency

Distributed workers maintain heartbeat leases. Crashed worker nodes release task locks automatically for zero-downtime recovery without duplicate execution.

3. Cloudflare R2 Artifact Vault

Agent outputs, generated code, diffs, and PDF receipts are written to S3-compatible R2 storage with zero egress fees and verifiable download links.

The Systems Behind the Governed Lifecycle

Beyond the execution core, CadreGPT ships the subsystems teams need to evaluate, observe, ground, guard, version, and publish agents inside one workspace.

Agent-to-Agent (A2A) Protocol

Agents invoke subordinate specialized agents with structured parameters, bounded recursion, and provenance tracing down the organizational chart.

Persistent Metric Time-Series Store

Historical metrics are persisted in PostgreSQL with time-series indexing, enabling period-over-period trend analysis, scoring, and attribution.

Automated On-Call Paging Alerting

The AlertService continuously evaluates BullMQ queue backlogs and task failure rates, automatically paging on-call engineers via webhooks and email.

Evaluation & Dry-Run Harness

Versioned eval datasets and a dry-run playground gate agent changes before promotion, scoring a candidate revision against the same prompt set as the current one.

Distributed Trace Store

Runs emit span-level telemetry (model calls, tool dispatch, tokens, latency) streamed over Server-Sent Events and persisted for post-execution review.

Vector Knowledge Retrieval

Workspace documents are embedded into pgvector-backed indexes and retrieved under strict per-workspace scoping to ground agent answers.

Governed Agent Memory

Cross-run memory is tenant-partitioned with configurable TTL and on-demand deletion, so retained context never leaks across workspaces.

Policy Guardrail Engine

Runtime policies enforce PII redaction, JSON output schemas, topic bounds, and per-run cost ceilings before an action executes.

Agent Lifecycle Management

Agent definitions are revisioned with diff, promote, and rollback, making every prompt or model change auditable and reversible.

MCP Tool Gateway

Model Context Protocol servers are registered and their tools discovered automatically, then handed to agents under the same approval gates.

Embed Token Gateway

Agents publish through hashed, origin-locked, rate-limited embed tokens that back both the chat widget and public API endpoints.

Handoff Inbox

Conversations route to a workspace-scoped human handoff inbox with claim, thread, and resolve states when an agent needs a person.

Workspace Analytics

Per-workspace reporting aggregates spend, run volume, success rates, and latency so the AI workforce is measurable.

Knowledge Base & FAQs

Architecture & Compliance FAQs

Deep-dive into our execution runtime, enterprise security isolation, and compliance standards.

01.What LLM models are supported?

CadreGPT runs on a managed multi-provider gateway: Google Gemini (2.5 Pro, 2.5 Flash), Anthropic Claude (3.7 Sonnet, 3.5 Haiku), OpenAI (GPT-4o, o3-mini), and any custom OpenAI-compatible inference endpoint or self-hosted model. Which model executes a task is decided by the platform routing matrix, operated by CadreGPT: each task is classified by profile — general execution, fast inference, complex reasoning, tool orchestration, long-context, or content — and dispatched to the model configured for that profile, with circuit-breaker failover.

02.Can I choose which model my agent uses?

No — model selection is platform-managed by design and is not a per-workspace or per-node setting. You describe the job; the router picks the model profile that fits it. That is what lets a workload move to a cheaper or a stronger model without you reconfiguring anything, and it is why a credit tracks the tokens a run actually consumes rather than a tier you have to shop for. You control the agent's prompt, tools, turn limit and cost ceiling.

03.How is data isolated between workspaces?

Data isolation is strictly enforced at both database and execution layers. Multi-tenant PostgreSQL schemas utilize workspace-scoped foreign keys and Row-Level Security (RLS) policies. Execution artifacts, memory vectors, and Redis queue streams are cryptographically partitioned with tenant-specific namespaces.

04.Are executions SOC 2 compliant?

Yes. CadreGPT architecture adheres to SOC 2 Type II trust security principles. All tool credentials and API keys are stored in an AES-256-GCM encrypted vault with step-up verification. Complete immutable audit logs, tamper-evident action receipts, and human-in-the-loop verification gates ensure full compliance across all autonomous agent actions.

05.How does BullMQ handle worker crashes and task retries?

BullMQ and Redis Redlock maintain active heartbeat leases on every running task. If a worker instance fails or becomes unresponsive, the heartbeat lease expires automatically, releasing the task lock for zero-downtime recovery and deterministic retry without data corruption or duplicate tool side-effects.

06.Where are agent outputs and generated artifacts stored?

Generated files, source code diffs, executive PDF reports, and execution logs are securely persisted to Cloudflare R2 object storage with zero egress fees. Access is managed via short-lived, signed URLs scoped to authorized workspace members.

07.How does CadreGPT keep agent knowledge and memory tenant-isolated?

Knowledge documents, embedding vectors, and agent memory are partitioned by workspace. Retrieval queries run scoped to the requesting workspace only, and memory entries carry an expiry you set or can be deleted on demand.

08.Can we govern and reverse agent changes?

Yes. Every change to an agent definition is stored as a revision. You can diff two revisions, promote a candidate, or roll back to a prior version, and each action is captured in the workspace audit trail.

09.How are guardrails enforced at runtime?

Guardrail policies run before execution and during tool dispatch: PII is blocked, structured output must match a defined JSON schema, agents stay within approved topics, and every run is bounded by a cost ceiling.

10.How do evaluation and dry-run work?

The dry-run playground executes an agent against a sample input so you can inspect its plan and output first. The evaluation harness runs the agent across a saved dataset and scores each result, so you can compare a candidate revision against the current one on identical prompts before promoting it.

Explore the Platform in Action

Ready to inspect the execution traces, streaming logs, and approval gates?

Launch Free Sandbox Workspace →