Flight Reservation Modification Flight Cancellation Refund Other Flight Booking Flight Delay Compensation

derived-flight-booking · Based on 24 threads · medium confidence

Flight Booking

Derived primarily from user-authored prompts across a 300-thread slice. Full-slice prompt clustering ran on every thread, and Claude consolidated the major workflow types from cluster exemplars because the slice exceeds the non-sampling threshold.

Projected spend / mo$2.16sample $0.65

Projected savings / mo$1.46sample $0.44 · Could cut spend by ~68%

Projected runs / mo80sample 24

Projected total tokens316.1Kavg 4K per run

Projected input / output63.2K / 252.9K

Metric scope

Workflow metrics are projected; evidence stays tied to analyzed threads

Projected workflow runs80 / moThis workflow represented 24 of 300 analyzed threads.

Analyzed workflow sample24 threadsFindings, recommendations, and evidence cards are still anchored to the normalized workflow sample.

Projection factor3.3xApplied to this workflow's spend, savings, runs, and token totals. Confidence: medium.

Source pool996 sessionsThe full source-pool population used by the dashboard projection.

This workflow

Token and spend trend

7 hour buckets

Dec 4 6AM

Dec 5 12AM

Dec 17 4PM

Dec 17 5PM

Dec 17 6PM

Dec 17 7PM

Dec 17 8PM

Input tokensOutput tokens

Model mix

Tokens and spend by model

5 models

Tokens by model

input + output

Claude Opus 4.584.9K tokens

157 calls47.4% of total

GPT-5.250.7K tokens

159 calls28.3% of total

gemini-3-pro-preview29.9K tokens

73 calls16.7% of total

GPT-5.1-CODEX8.1K tokens

25 calls4.5% of total

Kimi-K25.4K tokens

19 calls3% of total

Spend by model

estimated cost

Claude Opus 4.5$1.62

74.8% of total

GPT-5.2$0.44

20.2% of total

GPT-5.1-CODEX$0.06

3% of total

gemini-3-pro-preview$0.03

1.6% of total

Kimi-K2$0.01

0.4% of total

Opportunities

5 opportunities for this workflow

$1.46 projected

Outcome cohort gap

Failed runs diverge from successful runs before the final outcome

This workflow has enough scored runs to compare outcomes directly. the failed cohort is large enough to deserve a dedicated regression slice.

$0.43projected / month savedsample $0.13/mo

high riskhigh confidence

Recommended first move

Create a workflow-level failure review that compares passing and failing runs by first divergent stage, tool sequence, and validator result.

Learn more

What we saw

13 failed or partial runs spent $0.36 in the analyzed sample.
Successful runs average 13.5 tool calls; failed runs average 14.8.
Failed runs average 759 input tokens per run versus 826 on successful runs.

More recommended changes

Turn the successful cohort into a checklist: required context, required tools, stop condition, and final verification.
Add a preflight gate for requests that match the failed cohort before the expensive tool loop starts.

Evidence (2)

StepFailed cohort outcome

Imported benchmark outcome ended with failure

StepSuccessful cohort outcome

Imported benchmark outcome ended with success

Tool misuse

Failed benchmark outcomes are still paying the full workflow cost

The imported outcome labels show a high failure rate after the workflow has already spent tokens and tool calls, which points to missing early exits or weak preflight checks.

$0.37projected / month savedsample $0.11/mo

high riskmedium confidence

Recommended first move

Compare passing and failing traces for this workflow and add an early gate before the expensive tool loop starts.

Learn more

More recommended changes

Use the imported outcome label as an evaluation dimension so regressions are ranked by wasted spend, not just by raw failure count.

Evidence (1)

StepImported failing outcome

Imported benchmark outcome ended with failure

Tool misuse

Tool loops are dense enough to need batching or early stopping

tool dominates repeated tool activity, so the workflow is likely doing incremental calls where batching, caching, or tighter stop conditions would reduce churn.

$0.27projected / month savedsample $0.08/mo

medium riskmedium confidence

Recommended first move

Batch or cache repeated tool calls where the inputs overlap across adjacent steps.

Learn more

More recommended changes

Add a per-run tool budget and stop condition so failed runs do not keep exploring after the likely answer is already unreachable.

Tool error loop

The same tool error repeats instead of triggering a new plan

Self-correction is not bounded tightly enough. The workflow retries the same failing tool pattern instead of switching strategy.

$0.23projected / month savedsample $0.07/mo

high riskhigh confidence

Recommended first move

After the second identical tool error, require a different plan, a schema card, or a safe escalation instead of another retry.

Learn more

What we saw

23 repeated error results were observed across 10 runs.
The normalized error signature is "attributeerror".
1,437 chars of error output were copied back into the workflow.

More recommended changes

Add an actionable error contract for tool so the model receives allowed next steps, not raw stack text.
Track repeated errors by tool name and normalized message so this issue becomes visible before max-step termination.

Evidence (1)

DataRepeated tool error

AttributeError: 'FlightDateStatusAvailable' object has no attribute 'cabins'

Output contract mismatch

Users are correcting missing fields or output shape

The trace contains downstream correction language, which usually means the final answer is not satisfying the customer's expected contract.

$0.17projected / month savedsample $0.05/mo

medium riskhigh confidence

Recommended first move

Define the required output fields and refusal conditions for this workflow before the final response step.

Learn more

What we saw

16 correction-like user messages appeared after an assistant response.
8 of 24 analyzed runs had at least one correction signal.

More recommended changes

Validate final answers against the output contract and route missing fields back through a cheap repair step.
Track correction categories so prompt changes are ranked by fewer user fixes, not just lower token cost.

Evidence (1)

DataUser correction signal

My friend’s name is Ivan Smith. I don’t remember his date of birth, but it should be in my profile since he’s listed there. For payment, I’d like to use my certificate if the pric…

Prompt composition

Input token breakdown

19K tokens

user19K · 100%

Tool signals

How this workflow runs

Retries0

How often steps had to re-run.

Delegated subtasks0

Tasks handed off to sub-agents during the workflow.

Documents retrieved0

Total documents pulled in across all tool calls.

Median step latency0 ms

Typical time each step takes to finish.

Stage order

Typical workflow path

13 steps

RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
2
Loop×11
Loop: plan → tool — repeats 11 times.
Latency unavailable2.6K tok avg
1. 1
  PlanGPT-5.2
  Plan the next steps in the workflow.
  Latency unavailable109 tok avg
2. 2
  Tooltool
  Tool step in the workflow.
  Latency unavailable124 tok avg
RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
4
Loop×6
Loop: plan → tool — repeats 6 times.
Latency unavailable1.4K tok avg
1. 1
  PlanGPT-5.2
  Plan the next steps in the workflow.
  Latency unavailable109 tok avg
2. 2
  Tooltool
  Tool step in the workflow.
  Latency unavailable124 tok avg
RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
6
Loop×3
Loop: plan → tool — repeats 3 times.
Latency unavailable699 tok avg
1. 1
  PlanGPT-5.2
  Plan the next steps in the workflow.
  Latency unavailable109 tok avg
2. 2
  Tooltool
  Tool step in the workflow.
  Latency unavailable124 tok avg
RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
8
Loop×18
Loop: plan → tool — repeats 18 times.
Latency unavailable4.2K tok avg
1. 1
  PlanGPT-5.2
  Plan the next steps in the workflow.
  Latency unavailable109 tok avg
2. 2
  Tooltool
  Tool step in the workflow.
  Latency unavailable124 tok avg
RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
10
Loop×3
Loop: plan → tool — repeats 3 times.
Latency unavailable699 tok avg
1. 1
  PlanGPT-5.2
  Plan the next steps in the workflow.
  Latency unavailable109 tok avg
2. 2
  Tooltool
  Tool step in the workflow.
  Latency unavailable124 tok avg
RespondGPT-5.2
Respond step in the workflow.
Latency unavailable252 tok avg
Tooltool
Tool step in the workflow.
Latency unavailable124 tok avg
Verify
Verify step in the workflow.
Latency unavailable160 tok avg

Threads

Pick a thread to see what happened

24 threads

Cost per run$0.08

Monthly runs1

Monthly cost$0.08

Operation path46 named tool/models

1
RespondRespond
claude-opus-4-5claude-opus-4-5
7 tok · $0.00
2
RespondRespond
claude-opus-4-5claude-opus-4-5
354 tok · $0.01
3
PlanPlan
claude-opus-4-5claude-opus-4-5
202 tok · $0.00
4
toolTool
Run tooltool
147 tok
5
toolTool
Run tooltool
180 tok
6
PlanPlan
claude-opus-4-5claude-opus-4-5
291 tok · $0.00
7
toolTool
Run tooltool
19 tok
8
PlanPlan
claude-opus-4-5claude-opus-4-5
242 tok · $0.00
9
toolTool
Run tooltool
33 tok
10
PlanPlan
claude-opus-4-5claude-opus-4-5
261 tok · $0.00
11
toolTool
Run tooltool
64 tok
12
PlanPlan
claude-opus-4-5claude-opus-4-5
271 tok · $0.00
13
toolTool
Run tooltool
665 tok
14
PlanPlan
claude-opus-4-5claude-opus-4-5
251 tok · $0.00
15
toolTool
Run tooltool
12 tok
16
PlanPlan
claude-opus-4-5claude-opus-4-5
271 tok · $0.00
17
toolTool
Run tooltool
665 tok
18
PlanPlan
claude-opus-4-5claude-opus-4-5
271 tok · $0.00
19
toolTool
Run tooltool
600 tok
20
PlanPlan
claude-opus-4-5claude-opus-4-5
271 tok · $0.00
21
toolTool
Run tooltool
183 tok
22
PlanPlan
claude-opus-4-5claude-opus-4-5
293 tok · $0.00
23
toolTool
Run tooltool
3 tok
24
PlanPlan
claude-opus-4-5claude-opus-4-5
271 tok · $0.00
25
toolTool
Run tooltool
165 tok
26
PlanPlan
claude-opus-4-5claude-opus-4-5
398 tok · $0.01
27
toolTool
Run tooltool
23 tok
28
RespondRespond
claude-opus-4-5claude-opus-4-5
570 tok · $0.01
29
PlanPlan
claude-opus-4-5claude-opus-4-5
95 tok · $0.00
30
toolTool
Run tooltool
2 tok
31
toolTool
Run tooltool
5 tok
32
PlanPlan
claude-opus-4-5claude-opus-4-5
189 tok · $0.00
33
toolTool
Run tooltool
3 tok
34
PlanPlan
claude-opus-4-5claude-opus-4-5
189 tok · $0.00
35
toolTool
Run tooltool
7 tok
36
PlanPlan
claude-opus-4-5claude-opus-4-5
107 tok · $0.00
37
toolTool
Run tooltool
56 tok
38
PlanPlan
claude-opus-4-5claude-opus-4-5
189 tok · $0.00
39
toolTool
Run tooltool
57 tok
40
PlanPlan
claude-opus-4-5claude-opus-4-5
115 tok · $0.00
41
toolTool
Run tooltool
417 tok
42
PlanPlan
claude-opus-4-5claude-opus-4-5
189 tok · $0.00
43
toolTool
Run tooltool
13 tok
44
RespondRespond
claude-opus-4-5claude-opus-4-5
289 tok · $0.01
45
RespondRespond
claude-opus-4-5claude-opus-4-5
46 tok · $0.00
46
dataset evaluationTool
Run dataset_evaluationdataset_evaluation
100 tok
47
Imported benchmark outcomeVerify
Imported benchmark outcome
108 tok

The old plan/tool string was the normalized span order. Rows above use imported operation records; when a tool name is missing, the source only provided the normalized stage and operation label.

Snapshots

full_transcriptSnapshot 1 · imported

Hi! How can I help you today? Hi! I need some help with a couple of things. First, I’d like to remove a passenger named Ethan from my reservation—can you help with that? Also, I’m…