ThreadSync Architecture
The five-layer Agent Control Stack, decomposed for the technical buyer.
The Agent Control Stack is our public architecture. This page expands it into the seven implementation components we operate, with the comparison stance we hold against adjacent frameworks.
The 5 Public Layers, Decomposed
The public stack is five layers because that is the right cognitive load for a buyer. At the implementation level, two of those layers each decompose into two distinct components — yielding seven implementation pieces.
| Public Layer | Decomposes Into | What It Does |
|---|---|---|
| 1. Context | Context | Durable knowledge layer — operating doctrine, account history, policies. Memory that survives across runs. |
| 2. Skills & Playbooks | Skills / Playbooks | Named, versioned procedures with declared inputs and outputs. Composable into higher-order workflows. |
| 3. Integrations | Tools + Connectors | Tools (what the agent uses — email, DB, documents, APIs) and Connectors (where tools talk to — Gmail, Notion, Postgres, Slack, CRMs). The split matters at implementation: tools are scoped capabilities; connectors are auth-bound endpoints. |
| 4. Agents & Automation | Agents/Workers + Automation Hooks | Agents (specialized workers) and Hooks (triggers, checks, approvals, escalation rules). Public framing collapses these because the buyer cares about “what runs”; the technical implementation separates them. |
| 5. Governance | Governance | Permissions, audit trail, human approval, security boundaries. Cross-cutting across all other layers. |
This is implementation decomposition, not a parallel model. The five-layer stack remains the architecture. The seven components describe how the architecture is implemented.
What ThreadSync Provides at Each Layer
- ThreadSync Core — Postgres-backed substrate with structured models for customers, accounts, orders, tickets, transactions, communications.
- Connector implementations for CRM, ERP, email, data warehouse, financial systems, internal services.
- Contract-first skills with declared inputs, outputs, side effects.
- Versioned playbooks composing skills into business workflows.
- We operate this internally; we help clients build their own through advisory engagement.
- Connector patterns covering CRM, ERP, email, collaboration, and data-warehouse systems (Salesforce, SAP, NetSuite, Dynamics 365, ServiceNow, Workday, Snowflake, Gmail, Postgres, Notion, Slack, Google Drive). Implementation status varies; current availability discussed during engagement scoping.
- Custom API connectors via standard auth flows (OAuth, API keys with scoped roles, mTLS).
- Tool definitions follow the Anthropic / OpenAI tool-call schema; portable across providers.
- Magic Runtime — contract-driven execution with process isolation, default-deny permissions, and logs with write-once integrity controls.
- LLM Gateway — one governed access layer for five provider ecosystems. Policy-based routing, org policies, rate limits, budgets, cost tracking.
- Triggers via webhooks, schedules, event-driven rules, or human approval gates.
- Authentication at every layer boundary.
- Policy checks before action execution (default-deny).
- Immutable audit log of every request, response, and tool call.
- Human approval gates configurable at the policy layer.
- Cross-references: /security.html, /trust-center.html.
What ThreadSync is not replacing
ThreadSync is not a competitor to LangChain, AutoGen, CrewAI, the Claude Agent SDK, LangSmith, or your observability stack. We sit above them as the governance and operating frame.
| Layer | Common Open-Source / SDK | ThreadSync Adds |
|---|---|---|
| Context | Vector stores, RAG pipelines, custom DBs | Structured context substrate, normalized models, durable memory across runs |
| Skills & Playbooks | LangChain chains, AutoGen agent definitions, Claude Agent SDK skills | Contract-first declaration, versioning, audit, approval gates |
| Tools | Tool-call schemas (Anthropic, OpenAI, MCP) | Scoped capability binding, per-skill rate limits, budget caps |
| Connectors | Custom integrations | Scoped connector implementations with revocable auth, audit on every call |
| Agents | LangChain agents, AutoGen, CrewAI, custom orchestrators | Process isolation, default-deny permissions, immutable execution receipts |
| Hooks | Cron, Zapier, n8n, custom event buses | Policy-checked triggers, approval gates, escalation rules |
| Governance | LangSmith, Helicone, custom audit | Cross-cutting policy plane spanning all layers, with regulator-ready audit |
Bring your own framework. Build skills with the SDK you prefer. Use the agent orchestrator your team likes. ThreadSync provides the governance perimeter, the operating substrate, and the LLM access plane that make production deployment auditable.
What this page is not
- Not a sales pitch. For product positioning, see /platform.html.
- Not a security claim list. For security and trust posture, see /security.html and /trust-center.html.
- Not a competitive teardown. We do not compare ourselves to specific vendors by name in marketing; the categorical comparison above is the most we say publicly.