Developer Open-Sources Bottega Agent Orchestration Tool

In a blog post, Vincent Daubry announced he was open-sourcing Bottega, an internal agent-orchestration tool his team built and used in production, after shipping their 1,000th user story with it. He reports that for the past 8 months, 100% of the team's production code has been written by agents, with humans owning the plan and the review. Around the same time, Anthropic announced new Claude Code pricing; Daubry says Bottega's orchestration layer was easy to extend, so the team added support for Codex and OpenCode alongside Claude Code, which in turn lets Bottega run open-source models such as Kimi and DeepSeek. The practical benefit, per the post, is mixing models within a single task - for example, Claude Opus for planning, Claude Sonnet for implementation, Codex for code review, and an open-source model to manage the pull request.

Bottega runs as a web UI that is team-first (multi-user) and remote-first, designed to run on a shared server rather than a single developer's laptop, which Daubry says made sandboxing agents with skip-permission settings less risky and let non-developers at the company use it too. The pipeline separates planning, implementation plus unit tests, manual testing against scenarios defined in the plan, and PR management (an agent opens the PR, then iterates on CI failures and merge conflicts, including in response to GitHub review comments). Daubry positions Bottega alongside similar tools that emerged over the same period - Conductor (Melty Labs, YC S24), GasTown (Steve Yegge), gstack (Garry Tan), and GitHub's SpecKit - and says the overlap is confirmation the team converged on a workflow others are independently building toward.

Daubry's central argument is that the plan, not the model, is the centerpiece of the agentic development cycle: his team's earlier failure mode was treating the plan as a disposable, session-bound artifact rather than an enduring one, which produced large, hard-to-review PRs requiring heavy back-and-forth. Bottega's fix was to make the task, requirement, and technical spec persistent artifacts that live alongside the implementation, add a dedicated manual-testing step where the agent runs the scenarios defined in the plan before opening a PR, and add an adversarial review agent that checks the implementation strictly matches the plan. Separately, Daubry reports a two-week production bug in which Claude subprocesses silently defaulted to Sonnet 3.7 instead of a frontier model; output quality during that period stayed comparable to Opus 4.6, which he takes as evidence that a tight harness and rigorous process reduce sensitivity to model choice more than most teams assume.

Whether other teams report comparable production-scale agent usage and similar harness-over-model conclusions; whether open-source orchestration layers like Bottega gain adoption alongside Conductor, GasTown, gstack, and SpecKit; and how cost pressure from Anthropic's new Claude Code pricing shapes broader adoption of multi-provider model mixing and open-weight models such as Kimi and DeepSeek in production coding pipelines.

This is a single team's self-reported account, not an independently audited benchmark, so the 100%-agent-written and comparable-to-Opus-4.6 claims should be read as one practitioner's experience rather than a general result. That said, Daubry's observation that multiple teams converged on similar plan-first, multi-agent-role workflows around the same time (citing Conductor, GasTown, gstack, and SpecKit) suggests a real pattern in how production coding-agent workflows are being designed, independent of whether Bottega specifically gains adoption.