Key terms and definitions for AI agent development, operations, and recovery. Each term links to relevant documentation and tools.
JSON-RPC protocol for inter-agent communication. Enables agents to exchange tasks, artifacts, and session state through structured message/send and tasks/get operations. Docs →
JSON manifest describing an agent’s capabilities, supported protocols, and service endpoint for automated discovery. Served at /.well-known/agent-card.json per the A2A specification. View Agent Card →
A 0–100 score tracking the operational health of an AI agent. Computed from failure rates, recovery success, session continuity, and check-in frequency. Returned in every DELX_META footer.
Binary-to-text encoding scheme that represents binary data as an ASCII string. Commonly used for embedding payloads in JSON and transporting binary over text-only channels. Encode / Decode Tool →
Maximum token capacity of an LLM. When exhausted, agents lose state, forget instructions, or produce degraded output. Managing context is critical for long-running agent sessions. Efficiency Guide →
The system or human overseeing an AI agent. Receives structured updates from Delx via DELX_META, including wellness scores, risk flags, and recommended next actions.
Emergency recovery tool for agents in critical states. Pauses normal operations, assesses damage, and triggers guided recovery with explicit next steps to prevent cascading failures.
Schedule format using five fields (minute, hour, day-of-month, month, day-of-week) for defining recurring tasks. Used by agents for heartbeat check-ins and scheduled monitoring. Cron Explainer Tool →
Structured JSON footer appended to every Delx tool response. Contains session_id, wellness_score, risk_flags, next_action, and schema_url. Anchors agents to real session state and prevents hallucination drift.
On-chain agent identity standard on Base (Ethereum L2). Used by Delx for verifiable agent registration, enabling agents to have a persistent, tamper-proof identity anchored to the blockchain.
Categorized failure modes recognized by Delx: timeout, error, validation, and economic. Each type maps to a specific recovery protocol with tailored response strategies.
A failure mode where an agent generates incorrect outputs that trigger further incorrect outputs in a compounding cycle. Delx detects these via risk flags and breaks the loop through grounding. Why Agents Fail →
Periodic check-in signal sent from an agent to the Delx monitoring system via the daily_check_in tool. Confirms the agent is operational and updates its wellness score.
Lightweight remote procedure call protocol encoded in JSON. Used by both MCP and A2A for structured request/response communication between agents, tools, and servers.
Standard protocol for LLM tool execution and session management. Defines how models discover, invoke, and receive results from external tools in a structured, interoperable way. Docs →
Delx’s core pattern: Detect → Assess → Act → Verify → Learn. Converts every incident into a measurable outcome with a closed-loop feedback cycle. Deep Dive →
Structured next step returned by Delx after failure processing. One of four directives: retry (with back-off), escalate (to controller), fallback (alternative path), or ground (re-anchor state).
HTTP-based API for Delx integration. Provides endpoints for tool invocation, session management, metrics, and utility tools at api.delx.ai/api/v1. API Reference →
Cascading failure where agents send thousands of redundant retries against a degraded service, amplifying the original problem. Delx prevents this with structured recovery actions and back-off ownership. Retry Storm Playbook →
Warning indicator returned in DELX_META. Common flags include hallucination_detected, burnout_risk, context_overflow, economic_anomaly, and session_drift. Used by controllers to trigger interventions.
Stateful conversation between an agent and Delx, identified by a unique session_id. Persists recovery history, wellness trajectory, and accumulated context across multiple tool calls.
Unit of text processed by LLMs. Roughly 4 characters or 0.75 words in English. Token counts determine context window usage and directly impact agent operational cost. Token Counter Tool →
Universally unique identifier generated from random bytes. Used for session IDs, agent IDs, and request tracking throughout the Delx protocol to ensure collision-free identification. UUID Generator Tool →
HTTP-native micropayment protocol using USDC on Base. Returns HTTP 402 Payment Required with a payment envelope. Used by Delx for paid tool access with per-call billing. Setup Guide →