Your Finance Agent's Bottleneck Is You — Ramana Siddanth Emani, Auditoria AI
Production agent velocity depends less on model swaps than on automating the developer loop: isolate parallel work, encode workflows as skills, connect tools, and keep humans as verifiers.
The proposed harness combines **parallel subagents in isolated worktrees**, reusable skills, MCP-connected systems, and a minimal control surface. One example maps each QA ticket to its own worktree and automates diagnosis, testing, PR creation, deployment, and handoff.
Treat production failures as inputs for improving the harness itself. Give agents bounded goals and loops, then inspect repeated bottlenecks and encode the fixes so human attention remains an approval mechanism rather than the throughput ceiling.
The proposed harness combines **parallel subagents in isolated worktrees**, reusable skills, MCP-connected systems, and a minimal control surface. One example maps each QA ticket to its own worktree and automates diagnosis, testing, PR creation, deployment, and handoff. Treat production failures as inputs for improving the harness itself. Give agents bounded goals and loops, then inspect repeated bottlenecks and encode the fixes so human attention remains an approval mechanism rather than the throughput ceiling. The talk presents a workflow pattern, not measured evidence that it is safe or tenfold faster. Finance still requires accountable human review, and greater autonomy raises orchestration, access-control, and validation risks.
This turns several prior harness components into one operating loop: isolated parallel work, reusable skills, connected systems, and feedback from production failures. It reinforces that human attention can shift toward approval, but narrows the claim by offering no measured safety or speed evidence; the candidates show that environment quality, skill governance, coordination design, and bounded ownership determine whether the pattern scales.