Qualify a Stack Release
Compare a current and candidate coding-agent stack on private work and allocate workloads with evidence.
A Stack Release identifies the complete versioned coding-agent system. It can include the model and route, client or harness, instructions, skills, MCP configuration, hooks, workflow, permissions, reasoning budget, and required verification. A Harness Release can be one component of a Stack Release; it is not the comparison unit by itself.
For workspace owners
- Open Releases and freeze the current and candidate Stack Releases.
- Select a versioned private regression suite containing representative workload classes.
- Freeze the acceptance, cost, evidence-coverage, and release policy before execution.
- Select a customer-owned provider credential and run paired attempts in isolated Fullbeam-managed Daytona environments.
- Open Cost qualification and compare measured model and execution-attempt cost per accepted-task equivalent plus functional, CI, review, and regression guardrails.
- Allocate each supported workload: move to canary, keep the baseline, or continue evaluation.
- Record the policy-bound promote, scoped-promote, block, continue, or insufficient-evidence decision.
The required Fullbeam Qualification check enforces effective authorization on the exact PR commit. One continuously updated, aggregate-only PR comment explains the decision and links to the authenticated Decision Report. Use your coding-agent vendor, GitOps, MDM, or internal platform to deploy and roll back configuration. Fullbeam qualifies and records the decision; it does not perform the rollout.
Read the economics correctly
The product readout uses accepted-task equivalents: it sums each task's frozen repetition-aggregated pass rate, so majority and worst_case yield whole accepted-task counts while mean can yield fractional equivalents. It retains the measured cost of every valid attempt—including failed attempts—in the numerator. Raw scheduled-work cost is shown separately. If cost coverage is incomplete or no accepted-task equivalent exists, the economic result is unavailable rather than zero.
The current immutable Release Policy's economics gate still uses raw scheduled-work savings. Cost per accepted task is additional decision evidence; it does not silently replace or override the frozen policy result. A policy-bound cost-per-accepted-task gate requires a new policy and decision-contract version.
An annualized saving is also unavailable until the customer supplies an eligible workload volume or baseline spend, observation window, and other required assumptions. Fullbeam does not infer those inputs from token totals.
Runtime observation
Fullbeam starts the selected harness through its managed adapter and model gateway. At attempt start, it records the declared and materialized release, repository identity, adapter version, and applicable runtime policy. Observed runtime identity and downstream outcomes remain separate evidence with explicit coverage.
Fullbeam never silently changes local policy. Assigned/effective differences are evidence for the evaluation loop, not a deployment instruction.