CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 4 of 12 · 2 min read

Handing off the whole history or passing a brief: the cost crosses over around three specialists

Think of transferring a phone call to a colleague who can see your entire call history, versus one who only gets a two-line message describing what the customer needs. The first means nothing is lost. But every added colleague makes the message longer for all of them.

In plain words: a handoff passes control to another agent, which continues the conversation and sees the full history, while agent-as-tool keeps a manager in charge and calls the specialist with only a short brief, keeping just a summary on return -- so handoff's token cost grows roughly with the square of the number of specialists in the chain, while agent-as-tool's grows only linearly, and for assumed token sizes the two cross over at around three specialists.

A worked example. With a 2,000-token base context, 500 tokens added per turn, a 600-token brief and a 150-token summary: at one specialist handoff is cheaper (4,500 against 4,750 tokens). At two it's still cheaper (7,500 against 7,650). By three specialists agent-as-tool pulls ahead (10,700 against 11,000), and the gap widens fast -- at five specialists it's 17,250 against 19,500, and at eight, 28,200 against 36,000, a 22% saving.

The key lines. From code/mini-paradigms/paradigms.py:

# from code/mini-paradigms/paradigms.py
def handoff_tokens(k, base=2000, per_turn=500):
    """Triage then k specialists in sequence, each receives the full history (handoff)."""
    ctx, total = base, 0
    for _ in range(k + 1):                 # triage + k specialists
        total += ctx
        ctx += per_turn
    return total

Read it as: ctx grows by per_turn on every iteration and is added to total before it grows. This is exactly the arithmetic series that produces square-in-k growth. Each new specialist does not just add its own context. It inherits every previous specialist's growth too. That is the mechanical reason handoff cost accelerates. Agent-as-tool's as_tool_tokens resets each specialist to a fixed brief, so it does not accelerate.

Handoff, agent-as-tool

A handoff passes control to another agent, which continues the conversation and sees the history. Agent-as-tool keeps a manager in charge, which calls another agent like a function with a brief and keeps only a summary.

Example: the crossover at k=3 in this token model is not a universal fact about the two patterns -- it is a worked consequence of the specific assumed sizes (600-token brief, 150-token summary); a team should plug in their own numbers rather than treat "three" as a rule.