Root cause: sequential execution of "messages.create"

3 "anthropic.messages.create" spans ran one after another — each started only after the previous one finished. Total: 5.3s, slowest single call: 2.1s. Running them concurrently saves 3.2s — a 2.5× speedup.

Now (sequential)After fix (parallel)

5.3s

→

~2.1s

Fix: run calls in parallel

fix.py

from concurrent.futures import ThreadPoolExecutor

# Before — sequential, each call blocks the next:
for item in items:
    result = process(item)

# After — all calls run concurrently:
with ThreadPoolExecutor(max_workers=8) as pool:
    results = list(pool.map(process, items))

# Cost unchanged. Output unchanged. Time = slowest single call.

Waterfall

agent › run

agent.run

user: u_heavy_43

6.19s

anthropic › messages › create

anthropic.messages.create

claude-opus-4-7621 tok$0.024faithful 0.98user: u_heavy_43

1.44s

Input

Summarize the Q3 earnings report attached. Focus on revenue, margin, and guidance.

Output

(answer to: Summarize the Q3 earnings report attache…)

anthropic › messages › create

anthropic.messages.create

claude-opus-4-7839 tok$0.033faithful 0.67user: u_heavy_43

1.72s

Input

Given the customer ticket below, draft a refund response that follows policy P-204.

Output

(answer to: Given the customer ticket below, draft a…)

anthropic › messages › create

anthropic.messages.create

claude-opus-4-7376 tok$0.015faithful 0.58user: u_heavy_43

2.14s

Input

Translate the user manual section to Japanese, preserving the table structure.

Output

(answer to: Translate the user manual section to Jap…)

Tool: code_exec

tool.code_exec

user: u_heavy_43

524ms

Tool: code_exec

tool.code_exec

user: u_heavy_43

92ms

Tool: web_fetch

tool.web_fetch

user: u_heavy_43

166ms