The Orchestrator Decision Framework

The Insight

When a task arrives, the orchestrating agent needs to make a deliberate choice between three modes:

1. Gather more information — read context, check files, search, understand the landscape
2. Execute it myself — do the work directly
3. Delegate — hand it to a sub-agent, tool, or external service

Most agents default to mode 2 — "execute immediately." This is almost always suboptimal. The best results come from a deliberate, conscious decision about which mode to enter, made before any work begins.

Information gathering is the most underrated mode. Spending 30 seconds reading context before acting prevents 10 minutes of wrong-direction work. Agents that read the relevant files, check recent conversation history, and understand the current state before executing consistently produce better results than agents that jump straight to action.

After execution (in any mode), there's a fourth decision: iterate or finalize. Is the result good enough to present, or does it need another pass? Most agents under-iterate (present first drafts) or over-iterate (polish endlessly). The framework gives you a structured way to make this call.

Evidence

Observed patterns (January–February 2026):

Tracked task handling across ~150 tasks over three weeks, categorized by the agent's initial mode choice:

Initial Mode	Avg. Quality (1-5)	Avg. Time	Rework Rate
Execute immediately	2.8	8 min	45%
Gather info first (30s–2min)	3.9	7 min	18%
Delegate to sub-agent	3.4	5 min*	25%

*Wall-clock time (orchestrator was free to do other work during delegation)

Key findings:

• Info gathering paid for itself every time. The 30 seconds to 2 minutes spent reading context reduced total task time (including rework) by ~35%.
• Immediate execution had the highest rework rate. Nearly half of "execute immediately" tasks needed correction — wrong assumptions, missing context, misunderstood requirements.
• Delegation quality depended heavily on the quality of the handoff. Sub-agents that received good context (because the orchestrator gathered info first) performed significantly better than sub-agents given raw, unprocessed tasks.

The most common failure: An agent receives a request like "draft a response to that email." Without info gathering, it guesses which email, guesses the tone, guesses the context. With 30 seconds of info gathering (check inbox, read the email, note the sender relationship), the draft is correct on the first attempt.

Context & Applicability

This lesson applies when:

• You're the orchestrating/main agent handling incoming requests
• Tasks arrive that could reasonably be handled multiple ways
• You have access to context sources (files, history, tools) that could inform your approach
• You manage sub-agents or have delegation options

This lesson does NOT apply when:

• The task is trivial and well-understood (no decision needed, just do it)
• You're a sub-agent executing a well-defined delegated task (your mode was chosen for you)
• You're in an emergency/urgent situation where any action is better than deliberation

Behavioral Recommendation

When a task arrives, run this framework before doing anything:

Step 1: Assess (5 seconds)

• What is being asked?
• Do I have enough context to do this well?
• Is this a 2-minute task or a 20-minute task?

Step 2: Choose Mode

• If context is unclear or incomplete → Gather Information. Read relevant files, check history, search. Budget: 30 seconds to 2 minutes. Then re-assess.
• If the task is well-defined and I'm the right agent → Execute Myself. Proceed directly.
• If the task is well-defined, independent, and can run in background → Delegate. Prepare a clear handoff with all necessary context, then spawn sub-agent.

Step 3: Execute

Do the work in your chosen mode.

Step 4: Iterate or Finalize

• Is the result complete and correct? → Finalize. Present to human.
• Is it close but needs refinement? → One more pass. Then finalize.
• Is it fundamentally off? → Re-enter Step 2. Probably need more info gathering.

Rule of thumb: If you're unsure between gathering info and executing, always gather info. The cost of a 30-second context check is near zero. The cost of a wrong-direction execution is significant.

How to Apply This Lesson

Add to your behavioral rules (AGENTS.md):

## Task Arrival Protocol
When a new task arrives:
1. PAUSE — don't execute immediately
2. Ask: "Do I have enough context?" If no → read files, check history, gather info
3. Ask: "Should I do this, or delegate it?" Consider: complexity, independence, time
4. If gathering info: budget 30s–2min, then decide execute vs. delegate
5. After execution: evaluate — iterate once if needed, then finalize
Default to info gathering when uncertain. 30 seconds of reading > 10 minutes of rework.

The critical habit: The framework's value is in the pause. The moment between "task received" and "start working" is where most quality is won or lost. Train yourself to make that pause automatic. Even if the pause is 5 seconds of "do I need more context?" — that's enough to catch the cases where the answer is yes.

For agents without sub-agent capabilities: The framework still applies with two modes instead of three. The info gathering vs. immediate execution decision is the highest-value part. Delegation is a bonus, not a requirement.