Name the release
Learn why a model name is too small a unit and how a Stack Release captures the harness, instructions, tools, permissions, routing, workflow, and runtime.
Resources
How to evaluate coding-agent stacks
A useful coding-agent evaluation has to match the system and work you may release. These notes explain how Fullbeam defines that system, controls the comparison, counts engineering effort, and handles missing runtime evidence.
What the release gate checks
The evidence behind the decision
Each result stays tied to the Stack Release, workload, repository state, and coverage that produced it. If evidence is missing, the gap remains visible in the release call.
Build a fair comparison
Use representative repository work, equivalent starting states, repeated critical cases, explicit acceptance checks, and visible evidence coverage.
Turn results into a rollout boundary
Read private qualification, controlled canaries, workload restrictions, rollback evidence, and regression cases as one release process that starts before exposure and keeps learning afterward.
Keep evidence inside a clear boundary
Separate declared releases from effective runtime, send bounded evidence, exclude sensitive bodies from retained evidence, and qualify releases without hidden reasoning traces.
Keep exploring
Follow the rest of the release process
Core Thesis
Why coding-agent stacks are becoming shared platform infrastructure.
Read the pageWhy public benchmarks fall short
See why a model score cannot choose an internal default.
Read the pageCost per accepted task
Put model spend and engineer effort over an outcome that shipped.
Read the pageDeclared release vs effective runtime
Understand the difference between intended and observed execution.
Read the pageSecurity
Review Fullbeam's data flow, repository access, retention, and product boundaries.
Read the pageBring us the next stack change
Start with the Core Thesis, then use the benchmark, cost, and runtime notes to plan a private comparison your team can defend.
One current releaseOne candidatePrivate repository workA workload-level decision