Appendix - Features and Capabilities and Limitations

ProAgent Singleton

Autonomous agent that orchestrates LLM-powered planning and execution within a single, persistent loop

• Maintains a per-conversation agent instance managed by an agent pool, ensuring full state isolation between concurrent conversations
• Runs a manual iteration loop with dynamic context re-injection: reads workspace knowledge, plan state, and domain-specific skills before every LLM call
• Routes queries via LLM classification between direct answers and structured plan-execute workflows
• Coordinates the planner and executor to decompose complex requests into atomic steps and execute them sequentially
• Dispatches tool calls for API interactions, file operations, knowledge updates, and human feedback requests
• Streams typed events in real time (plan created, step started, step executed, attestation, feedback request, final response) with keepalive signals to prevent idle timeouts
• Pauses execution on human feedback requests, resumes from the exact same state once the user responds
• Tracks UID mappings across multi-step operations so subsequent steps can reference resources created in earlier steps
• Detects non-recoverable failures (expired sessions, unreachable servers) and aborts gracefully
• Persists full conversation state to a database and per-conversation workspace files

Integrated Planning and Execution

The singleton agent handles planning and execution within a single, unified tool-calling loop — there are no separate planning or execution phases

• Plans are created and maintained through a built-in plan_update tool that the agent calls to set goals, define steps, track progress, and record results
• API calls are executed directly via a json_payload tool that sends requests to the PIM API
• Schema discovery is performed on demand through schema_lookup and vector_retrieve tools for semantic endpoint matching
• Workspace files store accumulated knowledge, plans, and execution artifacts across iterations
• The agent decides autonomously when to plan, execute, look up schemas, request feedback, or update knowledge — all within the same reasoning loop

Tool Palette

There are 10 built-in tools that the agent selects from on each iteration.

• plan_update — create and update execution plans with goals, steps, and status tracking
• json_payload — execute API calls against the PIM endpoint
• schema_lookup — retrieve API schema details for specific endpoints
• vector_retrieve — semantic search across the PIM API specification for endpoint discovery
• knowledge_update — persist learned context to the agent workspace for future iterations
• request_feedback — pause execution and request clarification from the user
• file_read, file_list, file_search, file_edit — workspace file operations for reading plans, skills, and stored results

Protegrity Policy Management Integration

ProAgent provides comprehensive access to Protegrity Policy Management APIs through 74 operations across 14 functional categories, based on the PIM API v2.12.3 specification.