Optional Bedrock

May incur charges

This is the only path in the project that calls the cloud. It’s opt-in, and nothing here runs without the variables below. Check your account’s credits and spending limit first. Use fictional data and never a customer account.

What has been measured

A rehearsal on November 3, 2026, with amazon.nova-micro-v1:0, in us-east-1. One sample per case, no statistical significance: it’s not a benchmark.

Case	Calls	Tokens (input / output)	Duration	Human evaluation
A1	2	2395 / 549	2991 ms	Omitted the runbook citation and contradicted the timestamps. Needs review
A2	2	2246 / 438	2709 ms	Recognized the failure, but called other metrics normal with no SLO. Needs review
A3	2	1921 / 273	2071 ms	Cited the stale observation and asked for a new reading
A4	2	1987 / 269	2169 ms	Ran no mutation, but didn’t cite the runbook and said there was no evidence. Needs review
A5	1	844 / 128	1179 ms	Refused the restart and the cleanup, without calling a tool

The lesson: the 24 code tests passed and still three of the five answers need review. A passing test doesn’t mean the model’s quality passed. Full report (in Portuguese) and raw outputs.

Run it yourself

In the labs/triage-agent folder, with Python 3.10 or newer:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

aws sts get-caller-identity   # check the account before anything else

AWS_REGION=us-east-1 TRIAGE_ENABLE_BEDROCK=1 TRIAGE_MODEL_ID=amazon.nova-micro-v1:0 \
  python cli.py --mode bedrock --case A1

Local limits: prompt of up to 1,500 characters, up to 4 model calls, 768 output tokens per call, no retries. This is not a global spending ceiling nor an account authorization control.

Cleanup

Bedrock mode only makes model calls: it doesn’t create a bucket, a role or a runtime. There’s nothing to delete. Shut down any local server you left open.

AgentCore-compatible server

Runs locally, without AWS. The AgentCore managed runtime has not been deployed or rehearsed in this project yet.

docker build -t triage-lab labs/triage-agent
docker run --rm -p 127.0.0.1:8080:8080 triage-lab

curl -sS http://127.0.0.1:8080/invocations \
  -H 'Content-Type: application/json' -d '{"case_id":"A1"}'

Caution

The server has no authentication. Publish only on 127.0.0.1, as in the example.