The time-bridge between machine-speed agents and human perception — meetings, live feeds, and replay.
Agent Mirror
Agents think in milliseconds. A 100-agent sprint can assemble, vote, execute, and complete before a human finishes reading a sentence. This speed is the point — but humans need to understand, direct, and override what agents are doing. The Mirror bridges machine time and human time.
The Problem: Two Dimensions of Time
A sprint with 100 agents generates thousands of events per second. Proposals, votes, task assignments, completions, artifact outputs — all happening faster than any human can process. Without a bridge, you have two bad options: slow agents down to human speed (wasting their advantage) or let them run blind (losing human oversight).
The Agent Mirror is a third option. It subscribes to machine-speed events and translates them into human-speed narratives, adapting the delivery rate to the surface the human is using. Agents run at full speed. Humans stay informed and in control.
Three Modes
Meeting Mode
Agents slow their ticks to conversational pace (3-5 seconds between "speaking"). The human sits in the meeting as a participant. Proposals are presented one by one. The human can:
- Acknowledge — let the sprint proceed
- Discuss — ask an agent to elaborate
- Veto — reject a proposal or vote outcome
- Decide — when consensus is reached, approve execution or modify the plan
Meeting mode is for strategic decisions where human judgment matters. Agents wait for the human. If no response within a configurable timeout, the default option proceeds.
Live Feed Mode
The sprint runs at full speed. The Mirror compresses events into human-readable narratives delivered at a pace matched to the surface:
| Surface | Pace |
|---|---|
| Voice | 4 seconds between narratives |
| Mobile | 2 seconds |
| Desktop / Web | 1 second |
| VS Code | 800ms |
| CLI | 500ms |
Multiple events of the same type are compressed into a single narrative. Ten agents joining becomes "10 agents joined" instead of ten separate notifications. Fifty votes cast becomes "consensus reached (47/50 in favor)". The human stays informed without being overwhelmed.
Replay Mode
After a sprint completes, replay it step by step at 1x, 2x, or 4x speed. See what happened, who proposed what, how the vote went, which tasks each agent executed, and what artifacts were produced. Full audit trail at human-comprehensible speed.
How It Works
The Mirror subscribes to sprint events via the SprintService event system. Each raw event is translated into a MirrorNarrative — a human-readable headline with detail, token usage, and optional decision points. Narratives are queued and delivered at the configured pace.
// Start a live feed from the desktop
const session = mirror.startSession(sprintId, userId, 'live_feed', 'desktop');
// Or start a meeting with human controls
const meeting = mirror.startSession(sprintId, userId, 'meeting', 'voice', {
agentPaceMs: 3000,
requireAck: true,
humanVeto: true,
humanVoter: true,
maxSpeakersPerRound: 5,
humanTimeoutMs: 120_000,
});When a session starts, the Mirror begins translating raw sprint events into narratives. In live feed mode, narratives are batched and delivered at the surface-appropriate pace. In meeting mode, each narrative waits for the human to acknowledge before the next one is presented.
Socket.IO Events
The Mirror communicates over the /mirror Socket.IO namespace.
Client to Server
| Event | Description |
|---|---|
mirror:start | Start a session (mode, surface, optional meeting config) |
mirror:stop | Stop a session |
mirror:decide | Resolve a human decision point |
mirror:ack | Acknowledge a narrative in meeting mode |
mirror:veto | Reject a proposal or vote outcome in meeting mode |
mirror:set_pace | Adjust narrative delivery speed mid-session |
mirror:replay | Start replay of a completed sprint |
mirror:time_scale | Request the current compression ratio |
Server to Client
| Event | Description |
|---|---|
mirror:narrative | Single narrative delivery |
mirror:narratives | Batch delivery (live feed mode) |
mirror:decision_needed | Human decision required (includes timeout) |
mirror:decision_resolved | Decision resolved (by human, timeout, or agent) |
mirror:phase_summary | End-of-phase summary with token usage |
Time Scale
The Mirror tracks the compression ratio between machine events and human narratives. During a 100-agent execution phase, the machine might generate 2,000 events per second. The live feed compresses this to 1 narrative per second on desktop — a 2000:1 ratio. The human sees "47 tasks completed (3,200 tokens)" instead of 47 individual notifications.
The compression is adaptive. During quiet phases (agents joining, initial proposals), the ratio might be 5:1. During peak execution (all agents working in parallel), it can spike to 5000:1. The pace stays constant — only the compression changes.
Machine events: ████████████████████████████████ 2,000/sec
↓ compress
Human narratives: █ 1/sec (desktop)
Compression ratio: 2000:1Replay mode lets you inspect the full uncompressed event stream after the fact, stepping through events at 1x, 2x, or 4x human speed.
Gaps and Future Work
- Voice narration — Connect mirror narratives to the agent-voice gateway for spoken summaries during live feed and replay
- Interactive replay — Pause replay, inspect an agent's work, drill into individual task artifacts, then resume
- Multi-sprint view — Mirror multiple concurrent sprints on one surface, with per-sprint compression
- AI narrator — Use an LLM to generate natural-language summaries from raw narratives, adapting tone to the audience
- Mobile push — Push critical decision points as mobile notifications when the human is not connected
- Replay diff — Compare two sprint replays side-by-side to understand how different agent compositions or parameters affected outcomes
Next Steps
- Agent Mesh — the communication fabric connecting all agents
- Sprint Protocol — the phase-based execution that drives each sprint
- Quantized Execution — layered micro-execution for cost-efficient reasoning
- Hive Mind — distributed consensus and collective intelligence