The team’s response was to stop adapting a frontier-raised agent to a smaller model, and instead rebuild the agent from the ground up on open-weight, on-premises-compatible models only letting the constraints of the smaller model force better design from day one.

For the BA agent specifically, that meant restructuring a single monolithic document-generation pass into a phased, checkpointed workflow. The rebuilt version starts by setting only the input file and output location no upfront assumptions then extracts structured data from the legacy source only if it isn’t already available, and pulls simple counts of fields, program units, and blocks before the model does any generation work, deliberately without loading the full extracted dataset into context. Generation then proceeds section by section, with the largest, highest-risk sections written in small batches against an explicit zero-miss verification gate rather than in one continuous pass. Every phase writes to file; verification runs immediately after every write, checking counts against the extracted source of truth before the next phase is allowed to proceed. A final validation phase reconciles the completed output against the original source and produces a traceability report, and the workflow is only considered complete once both artefacts exist.

The next evolution, now underway, enforces this structure at the runtime level: an orchestrator issuing a fresh model call for each phase, with a hard checkpoint between hops, rather than relying on a single long-running session to hold the discipline. The trade-off is real each hop has less visibility into prior context, so cross-section consistency has to be carried through generated artefacts and structured data slices passed explicitly between hops, rather than through long conversational memory.