agent-steering

name	agent-steering
description	Background agent steering protocol — when to use write_agent vs launching fresh, run isolation guards, and cron exception rules. Use when user says "steer agent", "write_agent", "background agent", "run isolation", "inject context", "running agent", "agent communication", or any background agent interaction decision.

Agent Steering Skill

Canonical protocol for interacting with background agents via write_agent. Covers the steering pattern, run isolation guard, and cron exception.

Core Pattern

When an agent is running in the background (via task tool with mode="background") and new context arrives:

• Use write_agent to inject the update into the running session — don't kill and relaunch
• The agent will incorporate the new instructions while preserving its full context
• This avoids losing accumulated state (research, calculations, partial results)

Run Isolation Guard (Pipeline Agents)

⚠️ CRITICAL for multi-run agents (content pipeline, video production, client projects):

Only steer within the SAME run_id. If a new production run, video upload, or independent request arrives:

• ALWAYS launch a fresh agent instance via task tool
• NEVER inject a new run's context/assets into an agent processing a different run
• Cross-run contamination corrupts transcripts, research, deliverables, and publish targets

Examples

# ✅ CORRECT — follow-up within same run
write_agent(agent_id: "editor-abc", message: "YouTube URL is now available: https://...")

# ❌ WRONG — new video into existing session
write_agent(agent_id: "editor-abc", message: "New video uploaded: C:\vidpipe\new-video.mp4")
# ^ This contaminates the first run's context. Launch fresh instead.

Cron Exception (ABSOLUTE RULE)

Cron dispatches MUST ALWAYS launch fresh agents via task tool. Never use write_agent for cron-dispatched instances. The steering protocol is ONLY for interactive follow-ups within the same request/run.

See cron-dispatch skill for the full rationale.

When to Steer vs Launch Fresh

Scenario	Action
Follow-up in same run/request	write_agent ✅
New production run / new video	Launch fresh ✅
Cron-triggered job	Launch fresh ✅
Correction to ongoing work	write_agent ✅
Unrelated request, same agent type	Launch fresh ✅
Providing missing data agent asked for	write_agent ✅

Integration

• Complements agent-dispatch skill — that skill covers the orchestrator's decision; this skill covers the protocol once the decision is made
• Complements cron-dispatch skill — that skill covers why cron always gets fresh agents