Status: Accepted (March 2026) Context: As LLM capabilities evolve rapidly, we need a framework to decide where to invest engineering effort and where to hold still.
Decision
We classify every feature by its relationship to LLM progress, and allocate effort accordingly.| Category | Strategy | Investment | Examples |
|---|---|---|---|
| Orthogonal | Models getting smarter doesn’t diminish these — pure engineering/integration problems | Full investment | Connectors, credentials, OAuth, audit, RBAC, security, deployment |
| Tailwind | Models improving makes these better, not redundant — symbiotic relationship | Invest (benefits compound) | AI Connector Builder (smarter model = higher quality connector output) |
| Frozen | Already shipped, working well — but models are absorbing these capabilities | Maintain only, no new features | ReAct loop, DAG planning, RAG pipeline, Memory, Grounded Generation |
| Consider | Providers building natively at the platform level — high risk of redundancy | Deferred indefinitely | Multi-Agent orchestration, semantic memory, memory lifecycle |
Analysis
Why Connector Platform is fully orthogonal
Models will never natively:- Store and encrypt API credentials (AES-GCM)
- Manage OAuth flows (authorization page → callback → refresh token)
- Connect to a client’s Kingdee/金蝶 ERP database
- Push notifications to Lark/飞书 or WeCom/企微
- Enforce RBAC on who can use which connector
- Log every tool call for compliance auditing
Why AI Connector Builder is “tailwind” not “being absorbed”
The Builder Agent uses model intelligence to create managed, persistent Connector entities — stored in DB, reusable across agents, with credential management and audit trails. The model’s improving API understanding makes the Builder produce better connectors, not makes the Builder unnecessary. Analogy: Cursor uses Claude to write code. Claude getting smarter makes Cursor better, not redundant, because Cursor provides engineering value (project management, file organization, version control) that the model doesn’t replace.Why v0.1–v0.5 features are “frozen”
| Feature | What’s happening in the industry |
|---|---|
| ReAct loop | Models have native tool calling (OpenAI, Anthropic). The external reasoning loop adds less value as models internalize it. |
| DAG Planning | Model reasoning capabilities are improving rapidly. Complex task decomposition that needed external planners is becoming a single-shot capability. |
| Memory management | Context windows are growing fast (Gemini 2M+, Claude 200K+). The need for external window management, summarization, and compaction is shrinking. |
| RAG pipeline | Providers are building retrieval into their platforms (OpenAI file_search, Google NotebookLM, Gemini Search Grounding). For public knowledge, the traditional chunk-embed-retrieve pipeline is being replaced. |
| Grounded Generation | Models are getting better at citing sources natively. The 5-stage grounding pipeline we built adds diminishing value. |
Why Multi-Agent Orchestration was deferred
LLM providers are building orchestration natively:- OpenAI Swarm: Multi-agent framework with handoff protocols
- Anthropic Claude Code Teams: Leader/Worker agent pools with task graphs
- Google A2A (Agent-to-Agent): Inter-agent communication protocol
Why Semantic Memory and Memory Lifecycle were deferred
- Context windows are growing rapidly, reducing the need for cross-session memory retrieval
- Providers are adding native memory features (ChatGPT Memory, Claude Projects)
- The engineering cost of building a reliable memory system (TTL, importance scoring, semantic retrieval) is high relative to the shrinking gap it fills
Feature-Level Classification
Orthogonal (v0.6+)
| Feature | Version | Why orthogonal |
|---|---|---|
| Connector entity + CRUD | v0.6.1 | Enterprise integration, pure engineering |
| Per-user credentials (AES-GCM) | v0.6.2 | Security infrastructure |
| Confirmation Gate | v0.6.2 | Safety mechanism for write operations |
| Connector Export/Import/Fork | v0.7 | Distribution mechanism |
| OAuth 2.0 | v0.7 | Protocol implementation |
| MCP Server Export | v0.7 | Interoperability (depends on MCP adoption) |
| Database Connector | v0.8 | Direct DB access, connection pools |
| Message Push | v0.8 | Notification channels |
| RBAC | v0.8 | Access control, governance |
| Operation Audit Log | v0.8 | Compliance |
| Sandbox Hardening | v0.9 | Security isolation |
| Observability (OTel, circuit breaker) | v0.9 | Production operations |
| Connector Analytics | v0.9 | Usage tracking |
| Docker Compose | v0.9 | Deployment |
| Admin Dashboard | v1.0 | Management UI |
| Scheduled Jobs / Webhooks | v1.0 | Automation triggers |
| Batch Execution | v1.0 | Enterprise-scale processing |
| Embeddable Widget / iframe | v1.0 | Delivery mode |
| Enterprise Security | v1.0 | Compliance (encryption, IP whitelisting) |
Tailwind
| Feature | Version | Relationship |
|---|---|---|
| AI Connector Builder | v0.6.3 | Smarter models → better builder output |
| AI Connector Generation (OpenAPI) | v1.0 | Same — models understand API specs better → more accurate auto-generation |
Frozen (shipped, maintain only)
| Feature | Version | Status |
|---|---|---|
| ReAct Agent | v0.1 | Shipped, working |
| DAG Planning / Re-Planning | v0.1, v0.5 | Shipped, working |
| Memory (Window, Summary, Compact) | v0.2, v0.5 | Shipped, working |
| RAG pipeline (embedding, vector store, chunking, hybrid retrieval) | v0.5 | Shipped, working |
| Grounded Generation | v0.5 | Shipped, working |
| ContextGuard / Pinned Messages | v0.5 | Shipped, working |
Consider (deferred indefinitely)
| Feature | Original version | Reason deferred |
|---|---|---|
| Multi-Agent Orchestration | v1.0 | Providers building natively |
| Semantic Memory Store | Backlog | Context windows growing; providers adding native memory |
| Memory Lifecycle | Backlog | Same as above |
Implications
- Don’t go back to v0.5 features. Bug fixes yes, new capabilities no.
- Connector platform is the core investment. v0.6–v0.8 should receive the majority of engineering time.
- Enterprise engineering (RBAC, audit, security, deployment) is the moat. These are boring but defensible.
- Re-evaluate annually. If model progress stalls or a “frozen” feature turns out to still have significant gaps, reconsider.