Aurora — an agentic engineering environment

Context

Software projects move through steps: explore what exists, plan, implement, re-verify, ship. Standalone generative AI tools handle these steps in a single conversation window — and lose the project state every session.

Problem

Without persistent state or separated roles, the AI re-explores what it already explored, reinvents decisions already made, and reviews its own work. The result is not industrialisable.

Constraints

• Project state must survive sessions.
• Structuring decisions must be recorded, not reinvented.
• Every delivery must be reviewed from a different point of view.
• Capabilities (docs research, audits, deployment) are added without bloating the base prompt.

My role

Designing the architecture (orchestration, specialisation, state), building the whole setup, keeping a readable and open source configuration.

Approach

An orchestrator agent that delegates to specialised sub-agents — architecture, interface, security, review, testing, fast execution — rather than a single assistant doing everything. Project state lives in documents versioned inside the repository: status, plan, decisions, warnings. Every cycle ends with an adversarial review and a verification pass (build, tests) before shipping.

Architecture

• Dozens of specialised agents (architecture, UX, security, review, testing, execution, research…), one per role.
• On-demand injectable skills (code review, verification, commit, deployment…).
• Several MCP servers for external capabilities (docs, browser, mobile…).
• A documentary state layer (« Agent State Layer »): status, plan, decisions and warnings versioned in the repository.
• Model routing with automatic fallback: the right model for each agent, with no single-vendor dependency.

Key decisions

• State in the repository, not in the context window: every session starts from documents, never from conversation memory.
• Review is a distinct agent: whoever writes does not validate — adversarial review and verification run in parallel before any merge.
• Capabilities on demand: skills load when they serve; the base context stays light.
• Open source by default: the configuration is public (MIT) and documented.

Challenges

The hard part is not the number of agents but their boundaries: who decides what, and how to avoid loops. The documentary state layer is the answer — a clear write contract between orchestrator and sub-agents.

Result

An agentic engineering environment in daily use on real projects, open source under the MIT license on GitHub (himuraxp/opencode-config).

What this project demonstrates

Industrialising AI in engineering workflows: specialised, governed agents with persistent state and systematic review — no autonomy without guardrails.