A rough ordering of the shapes by how much the developer controls the path: sequential, conditional and parallel workflows give the most control and the least autonomy; plan-based orchestration gives medium autonomy and still high control; handoff gives medium autonomy and medium control for little complexity; an open conversation between agents gives the most autonomy and the least control. A common rule is to use workflows for well-defined repeatable processes, plan-based orchestration where planning is complex, handoff where domain expertise decides, and conversation for exploratory work. Production systems are usually hybrids: workflows for the predictable parts and autonomous patterns where flexibility matters.
Setup: tickets need one or two of billing, tech and legal; a ticket is correct only if every needed part is correct. 20,000 tickets per shape, seed 0.
Shape
Accuracy
Single-domain
Two-domain
Calls
Tokens
Critical path
single agent
0.63
0.69
0.48
1.0
1,300
1.0
route
0.53
0.76
0.00
2.0
2,270
2.0
orchestrator-workers
0.71
0.76
0.59
3.2
3,459
2.9
handoff
0.74
0.81
0.58
2.3
2,405
2.3
orchestrator plus critic
0.80
0.84
0.71
5.0
5,671
4.4
Accuracy and tokens per shape; accuracy as the share of two-domain tickets rises; accuracy of the critic loop by critic quality.
Accuracy by the share of two-domain tickets:
Share
0.0
0.3
0.6
1.0
single
0.70
0.63
0.57
0.49
route
0.76
0.53
0.30
0.00
orchestrator-workers
0.76
0.71
0.66
0.59
handoff
0.81
0.74
0.66
0.57
orchestrator plus critic
0.84
0.80
0.77
0.72
Critic quality (orchestrator plus critic; the plain orchestrator scores 0.71):
Detects wrong drafts
Rejects correct drafts 0%
10%
30%
Calls
0.0
0.71
0.69
0.65
4.4 to 5.2
0.5
0.78
0.77
0.74
4.6 to 5.5
0.9
0.83
0.84
0.82
4.8 to 5.8
Handoff under stress (misrouted specialists notice 90 percent of the time, and the confusion rate rises): with confusion 0.7, a hop limit of 2 cut off 3.3 percent of tickets and scored 0.737; a limit of 4 cut off 0.2 percent and scored 0.762; a limit of 8 cut off none. Revisiting an agent (a loop) happened in 0.6 percent of tickets at most.
Reading it:
Match the shape to the task mix. Routing is the best value for single-domain work and unusable for work that needs two domains. Orchestrator-workers is robust to the mix because it decomposes.
Handoff recovers misroutes (0.81 against 0.76 for plain routing on single-domain tickets) at lower cost than an orchestrator, because a misrouted specialist can pass the ticket on, while a router cannot take its decision back.
The orchestrator's accuracy is capped by its decomposition (0.90 per domain), so a weak orchestrator limits the whole system. This is the single-point-of-failure cost of a central coordinator.
The critic buys the most accuracy and the most latency (path 4.4 against 2.9). It helps even when imperfect (detects 0.5, rejects 0.1 wrongly: 0.77), but it can hurt, and a critic that agrees with everything is pure overhead.
Loops are rare in this model, so the hop limit protects the worst case rather than the average. Real systems loop more when handoff criteria are ambiguous, which this model does not capture.
Choosing a shape
Fixed process, known steps: a workflow (sequential, conditional or parallel). Lowest cost and highest predictability.
One clear domain per request: route to a specialist.
Requests that must be decomposed at run time: orchestrator-workers; give the orchestrator the best model and typed results (module 8 of Level 7).
Specialists that know best when a request is not theirs: handoff with a hop limit.
A check that is cheaper than the error: add a critic, and measure its detection and false-rejection rates first.
Exploration with no fixed path: conversation-driven shapes; budget them by message count.
Compose: a node can be a sub-system, so use a workflow inside an orchestrator, or a critic inside a worker.