Bounded success
99.4%exact-derivedExact terminal success within the maximum attempts.
[3]Portfolio 18 · bounded stochastic workflow
AI agent workflow economics simulator
Reconcile every expected node execution, token, tool call, and human-review minute; solve independent branches exactly and reserve seeded simulation for tail latency or correlated events.
Decision this answers: What does one successful workflow cost, which node drives it, and how much reliability and p95 latency come from a hard-bounded retry policy?
Scenario
Decision summary
Expected workflow cost
$0.40exact-derivedModels, tools, and expected human review.
[4][2]Cost per successful outcome
$0.41exact-derivedExpected cost divided by bounded terminal success.
[4]Simulated p95 latency
7.40model estimateSeed 42, 20,000 runs; success standard error 0.001.
[5]The tables below are the accessible source of truth; bar lengths never carry identity alone.
Workflow execution ledger
| Candidate | Expected executions | Token meter | External meter | Termination | Method | Evidence |
|---|---|---|---|---|---|---|
| Planner | 1.000 | Input + output | None | Acyclic | Exact | exact-derived[3] |
| Tool loop | 1.212 | Per attempt | Tool call per attempt | 3 attempts | Finite geometric | exact-derived[2] |
| Human exception | 0.060 | None | 4 min at 6% reach | Terminal escalation | Exact expectation | model estimate[5] |
Method boundary
| Candidate | Use when | Correlation | Uncertainty | Reproducibility | Evidence |
|---|---|---|---|---|---|
| Exact expectation | Acyclic independent branches or bounded retries | Independence declared | None | Closed form | exact-derived[3] |
| Monte Carlo | Tail percentiles or shared events | Explicit shared shock | Standard error | Seed + run count | model estimate[5] |
Official surface versus this site
Building effective agents
Use the official surface to confirm native prices, counting rules, limits, and product eligibility.
This offline workbench
The official guidance describes workflow and agent patterns. This page adds an auditable economic graph with termination bounds, exact expectation, reproducible uncertainty, and cost per successful outcome.
Sources and methodology
Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.
[1] Building effective agents
- Authority
- Anthropic
- Native identifier
workflow and agent patterns- Unit
- workflow topology
- Evidence
- exact-derived
- As of
- 2026-08-01
- Retrieved
- 2026-08-01
- Freshness
- reference
- Normalization
- The patterns define graph shapes; the site requires explicit reach, termination, and measured meters.
[2] Tool use overview
- Authority
- Anthropic
- Native identifier
tool use- Unit
- tool calls
- Evidence
- measured
- As of
- 2026-08-01
- Retrieved
- 2026-08-01
- Freshness
- reference
- Normalization
- Tool invocations remain separate cost-bearing workflow nodes.
[3] ADK workflow agents
- Authority
- Native identifier
sequential, parallel, and loop agents- Unit
- workflow control
- Evidence
- exact-derived
- As of
- 2026-08-01
- Retrieved
- 2026-08-01
- Freshness
- reference
- Normalization
- Loop-agent topology is accepted only with an explicit local maximum attempt bound.
[4] OpenAI Agents SDK usage
- Authority
- OpenAI
- Native identifier
RunResult usage- Unit
- request and token usage
- Evidence
- measured
- As of
- 2026-08-01
- Retrieved
- 2026-08-01
- Freshness
- reference
- Normalization
- Usage records populate node meters; no live request or telemetry is sent by this page.
[5] OpenAI Agents SDK tracing
- Authority
- OpenAI
- Native identifier
traces and spans- Unit
- node latency and reach
- Evidence
- measured
- As of
- 2026-08-01
- Retrieved
- 2026-08-01
- Freshness
- reference
- Normalization
- Compatible exported measurements inform reaches and latencies; the calculator itself remains offline.
Limits and non-claims
- Branch independence is an explicit model choice; shared failures and latency shocks require simulation.
- Averages do not prove p95 latency. The tail result is labeled with seed, run count, and sampling error.
- Success probability, expected cost, cost per success, and tail latency remain separate decision metrics.