We help your teams bring AI agents into their software development cycle: shared context, quality rules, tests, automation, determinism, an isolated environment and human review. The team defines the framework and learns to evolve it, keeping up with the state of the art.
Understanding context, respecting the architecture, testing and documenting are also part of development. We help your team make these expectations explicit and build them into the agents’ working environment.
Océanienne de Services Business
socraft helped the development team at Océanienne de Services Business agree on shared practices, then translate them into a harness around Claude Code. The framework covers documentation, testing strategy and where LLMs intervene, with checks throughout the cycle and human oversight.
Documentation, testing strategy and code review are discussed and written down with the team.
This framework connects instructions and automated checks to the development cycle.
A sandbox, or isolated environment, limits what the agent can touch.
Useful information is organised and shared with the team and, where needed, the agents.
The team defines what the agent may do on its own and what a developer must review.
The team agrees on its development rules and review points.
Rules become instructions and checks around the agent.
We work with the team on its real tasks to adapt the framework to its practices.
The team learns to evolve its rules and harness.
socraft · Internal project
We built a minimal harness around Pi. Adapted to our practices, it connects to several inference and model providers. It runs in an isolated environment that limits its scope of action, and evolves with us.
socraft · Internal project
A Slack bot built on Hermes and GBrain gathers knowledge about our environment and makes it available to the team.
At OSB, the framework was built around Claude Code. The approach applies to other agents: what matters is the team’s practices and the harness that turns them into instructions and checks.
The team decides: what the agent may do on its own and what a developer must review. A sandbox, or isolated environment, limits what it can touch.
A framework around the agent that connects the team’s instructions and automated checks to the development cycle: documentation, testing strategy, code review.
The team first agrees on its practices, then we build them into the harness and work with the team on its real tasks, until it evolves its framework on its own.
Tell us which agents your team uses, how it works and where it needs a clearer framework.
Book a free AI diagnostic