Use this as a field report when designing a GPT-6 Astra evaluation plan; validate its findings against our own engineering tasks.

Authorised code access does not remove risk: compare proposed patches with repository policy, test results and owners before merging.

Follow the end-to-end integration guide to add email summarisation, reply drafting, calendar scheduling and downstream workflow triggers.

Learned gist tokens warrant a full workload test: record task quality, TTFT, QPS, GPU allocation and end-to-end latency against prompts.

Long-horizon tasks expose failure modes short benchmarks miss; add original engineering work and inspect progress, recovery and changes.