Part 5 of 12 · 3 min read
What the example found
Crash simulation across three structures, and the handoff-versus-agent-as-tool token model.
How to read the crash-simulation table: each row is a different per-step crash probability; each paradigm's cell shows model calls per run, duplicate side effects per run, and the share of runs that finished within 6 attempts.
| Crash chance per step | Loop: calls / duplicates / finished | Graph + checkpoint | Graph + checkpoint + idempotency key |
|---|---|---|---|
| 0% | 6.0 / 0.00 / 100% | 6.0 / 0.00 / 100% | 6.0 / 0.00 / 100% |
| 5% | 6.9 / 0.17 / 99.95% | 6.3 / 0.11 / 100% | 6.3 / 0.00 / 100% |
| 15% | 9.1 / 0.50 / 96.9% | 6.9 / 0.34 / 99.7% | 6.9 / 0.00 / 99.7% |
| 30% | 11.6 / 0.56 / 67.5% | 8.0 / 0.81 / 94.5% | 8.0 / 0.00 / 94.5% |
What to notice: the graph shows more raw duplicates per run at 30% (0.81) than the loop (0.56). That looks backwards until you notice the graph also finishes far more runs (94.5% against 67.5%). It executes more steps overall simply by surviving more often. So duplicates have to be read next to the finish rate, never alone.
How to read the handoff-versus-agent-as-tool table: each column is a different number of specialists in the chain; each row is total input tokens under the assumed sizes (2,000-token base, 500 per turn, 600-token brief, 150-token summary).
| Specialists k | 1 | 2 | 3 | 5 | 8 |
|---|---|---|---|---|---|
| Handoff | 4,500 | 7,500 | 11,000 | 19,500 | 36,000 |
| Agent-as-tool | 4,750 | 7,650 | 10,700 | 17,250 | 28,200 |
What to notice: handoff's lead at k=1 and k=2 (250 and 150 tokens cheaper) is small next to agent-as-tool's lead at k=8 (7,800 tokens cheaper). The two patterns are not symmetric opposites. Agent-as-tool's linear growth makes its advantage compound exactly where the handoff pattern's quadratic growth starts to hurt most.
{{svg:paradigms|A loop, a checkpointed graph, and handoffs differ in where state lives and what happens at a crash.}} {{svg:results|Model calls per run and duplicate side effects per run, for three structures.}}
One published assessment says CrewAI and the OpenAI Agents SDK suit early prototypes while LangGraph and AutoGen suit scalable production. It will date. A capability checklist to ask of any framework, with the right-hand column showing what to check in current documentation rather than trust a stable answer:
| Capability | Why it matters | What to check |
|---|---|---|
| State and persistence | resume after a crash; memory across turns | where state is stored, how it is keyed, how old state is pruned |
| Pause and resume | approval before irreversible steps | is a pause a first-class feature or a workaround |
| Control flow | loops, branches, parallel work | explicit graph, or the model decides |
| Multi-agent | handoffs, delegation | context passed, who answers the user |
| Tools and protocols | MCP, function schemas | validation, parallel execution, error format |
| Guardrails and permissions | stop bad inputs, outputs, calls | run in parallel or before; deny rules |
| Observability | debug and evaluate | traces, cost and token reports, export |
| Provider portability | avoid lock-in | non-native models supported, and to what depth |
| Deployment | where it runs | library, hosted service, languages |
Guardrail, trace
A guardrail is a check on inputs or outputs that can stop a run. A trace is a recorded tree of model calls and tool calls for one run.
Example: as documented facts rather than this module's own measurements -- the Claude Agent SDK reports total_cost_usd, usage and turn count on every result and caps turns and spend; LangGraph documents checkpointers, threads and human-in-the-loop as one persistence mechanism; the OpenAI Agents SDK lists handoffs and guardrails as primitives and, by default, uses OpenAI's Responses API with adapters for other providers. Whether each is good enough for your case is for you to test, not for this checklist to answer.