Part 3 of 12 · 2 min read
An orchestrator's accuracy is capped by how well it decomposes the task
A coordinator that misses a needed specialist dooms the ticket before any specialist even gets involved.
Think of a project manager who splits a job among specialists. If the manager fails to notice the job actually needs an electrician as well as a plumber, no amount of plumbing skill fixes the part nobody was asked to do.
In plain words: the orchestrator decomposes the task, dispatches parts to workers and combines the results. A worker executes one part. But the orchestrator's own decomposition step is itself an imperfect call: it catches each truly-needed domain only some of the time. A ticket is doomed the moment the orchestrator fails to even recognise it needs a given domain, regardless of how accurate the eventual specialist would have been.
A worked example. With decomposition accuracy at 0.90 per domain (independently for each), orchestrator-workers scores 0.76 accuracy on single-domain tickets and 0.59 on two-domain tickets. That is lower than handoff's 0.81 and 0.58 respectively on the same splits. Every domain the orchestrator fails to notice is a domain no worker ever even attempts, while handoff lets a specialist who gets the wrong ticket notice and pass it along instead.
The key lines. From code/mini-shapes/shapes.py:
# from code/mini-shapes/shapes.py
def orchestrate(r, p, need, critic=False):
found = [d for d in need if r.random() < p.decomp_acc]
extra = [d for d in DOMAINS if d not in need and r.random() < p.spurious / 2]
parts = found + extra
calls, tokens, rounds_max, ok = 1, p.system + p.ticket, 0, len(found) == len(need)
Read it as: ok is initialised to len(found) == len(need), computed from found alone, before any worker has done anything. The ticket's fate is partly sealed at the decomposition step. A worker executing its one assigned part perfectly cannot rescue a domain that found never included in the first place.
Orchestrator, worker, decomposition
The orchestrator decomposes the task, dispatches parts and combines the results. A worker executes one part. Decomposition is splitting a task into parts that different agents handle.
Example: a decomposition accuracy of 0.90 per domain sounds high, but on a two-domain ticket it means both domains have to be independently caught. That is why orchestrator-workers' two-domain accuracy (0.59) sits well below its single-domain accuracy (0.76). The single point of decomposition is a single point of failure.