Compact Status Updates

Compact Status Updates (Micro-Telemetry Protocol)

Overview

When executing multi-step autonomous workflows, unoptimized agents emit lengthy conversational monologues before and after every tool call ("I will now search the directory for the database configuration. After that, I will inspect the schema to see if the column exists...").

These conversational narrations burn 150 to 300 output tokens per tool call. In a 20-step task, narrative chatter consumes 4,000+ tokens of pure transcript noise without advancing the actual codebase.

The Micro-Telemetry Protocol replaces conversational monologues with structured, 2-to-5 word tool metadata fields (toolAction, toolSummary) that feed directly into IDE progress bars and UI spinners with zero transcript bloat.

Conversational Monologue vs. Micro-Telemetry

┌─────────────────────────────────────────────────────────────┐
│                 Status Update Token Impact                  │
│                                                             │
│  Conversational Progress Monologue (165 Tokens):            │
│  "I am now going to invoke the grep tool to search for all  │
│   instances of 'jwt.verify' across the repository to see    │
│   how tokens are currently being validated in the routes..."│
│  ↳ 165 output tokens, slow streaming latency                │
│                                                             │
│  Micro-Telemetry Parameters (12 Tokens - 92.7% Reduction):  │
│  "toolAction": "Searching JWT Verification Instances",      │
│  "toolSummary": "JWT Verification Search"                   │
│  ↳ 12 tokens, instant UI progress bar update, zero noise    │
└─────────────────────────────────────────────────────────────┘

The Master Telemetry Parameter Standard

Every tool call in modern agent frameworks (Antigravity IDE, Claude Code) includes standardized UI telemetry arguments:

{
  "toolSummary": "Brief 2-4 word noun phrase describing the tool call (e.g. 'Directory Search')",
  "toolAction": "Brief 2-5 word active verb phrase (e.g. 'Searching Database Migrations')"
}

High-Efficiency Telemetry Examples:

End-of-Turn Executive Status Broadcast

When an agent completes a multi-step sequence, emit a single 1-to-2 sentence executive status update rather than re-listing every individual tool executed:

<!-- OPTIMAL END-OF-TURN BROADCAST -->
✅ Refactored `auth.ts` to validate JWT expiration against Redis revocation list. All 14 unit tests passing.

Benchmark Comparison

Evaluation across a 25-step autonomous feature implementation:

Agent Operational Directive

MANDATORY: Agents must NEVER write conversational narration paragraphs in chat prior to invoking a tool. Supply concise, capitalized 2-to-5 word phrases in toolAction and toolSummary, letting the UI render native progress states.