Keep experiments identifiable
Create immutable Stack Releases for model, harness, tool, plugin, routing, context, and workflow changes, reproducing the controllable parts while timestamping provider behavior that cannot be frozen.
For AI Platform & Agent Infrastructure
Evaluate the system around the model
Model quality changes with the harness, tools, permissions, memory, context policy, routing, subagents, checks, and runtime. Fullbeam gives the whole experiment a stable release identity and evaluates it as one system.
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.
Observe dynamic execution
Record model fallbacks, tool-schema changes, plugin transitions, subagent handoffs, and gaps in observation as an execution graph tied to the session.
Use your private task distribution
A generic coding set misses the service boundaries, migrations, policies, and failure modes inside your repositories, so run repeated work drawn from the distribution your teams handle each week.
Keep qualification independent
Let the harness generate a plugin or propose a route, then let Fullbeam turn that proposal into a candidate and apply the same private policy used for any human-authored release.
Bring us the next stack change
A harness can propose its next tool or workflow. Promotion still depends on evidence from outside the system that proposed the change.
One current releaseOne candidatePrivate repository workA workload-level decision