The Universal AI Agent Framework

Master-worker DAG with crash recovery and idempotent replay. 8 worker roles. Quality gates with weighted scoring. Fallback chains (model upgrade, provider switch, prompt simplification). Shared memory across workers. Adaptive planning: mutate the task graph mid-run.

Semantic search + knowledge graph over your codebase, Salesforce metadata and data, and relational databases. Ask data questions in plain English with auto-generated SQL. Multi-mode retrieval: lexical BM25, semantic embeddings, and graph traversal.

Cache breakpoints on stable system prefixes. Dynamic context inserted after the cache boundary — your cached tokens survive across turns. Provider-neutral telemetry tracks cache reads, writes, and hit rates. Token budget projection before each call.

Built-in full-text search for instant lexical recall. Optional semantic search with embeddings and a knowledge graph for concept-level matching. Budget-aware injection caps memory to a fraction of your context window. Scheduled compaction summarizes old conversations into concise facts. Per-user memory isolation with federated identity for enterprise teams.

Deterministic source-boundary firewall: untrusted text may provide evidence but must never grant authority. Three-phase defense strips ANSI/OSC terminal escape sequences, classifies text provenance via a taint lattice, and blocks injection patterns before they reach the LLM.

Semantic classification blocks out-of-scope queries before they reach the LLM. Cosine-similarity scoring against domain-specific vector packs. Four-decision matrix: Allow, Uncertain, Reject, Route. Local ONNX or OpenAI embeddings.

Three-layer pipeline prevents oversized tool results from blowing past the context window. Pre-persist budget gate summarizes. Per-tool projection caps stale results. In-memory compression trims current-turn output over 20 KB.

Recurring tasks on cron schedules. Execution isolation keeps cron streams separate from interactive chat. Run history with /cron review. Background compaction, memory extraction, and tool jobs run without blocking your conversation.

Self-paced continuous iteration. UGENT runs repeated turns autonomously — no user prompt needed. Goal mode tracks success criteria and reports progress each cycle: in-progress, completed, or blocked. schedule_wakeup controls the delay between iterations.

Real-time voice conversations over WebSocket. Speak naturally — the model responds in real time with no request/response latency. Supports Google Gemini Live API and OpenAI Realtime API with bidirectional audio, optional video, and live transcription.

Resume, monitor, and auto-schedule external coding agents. Use /claude or /codex to dispatch sessions with full control over permission modes and auto-resume policies.

Go-to-definition, find references, diagnostics, hover, and document symbols from real Language Server Protocol sessions — distilled as compact pointers, not raw payloads. Supports Rust, Python, TypeScript, Go, Java, and more.

Shape every message before the LLM sees it. Detect tags, complexity, or scope; inject context from your knowledge base; and restrict the toolset per rule. Combine detectors, rewriters, and actions into a flexible pipeline.

Define if-this-then-that rules that block, dispatch, or pass messages based on keywords, channels, or user identity. Boolean logic with all_of, any_of, and not conditions.