AIGo

AIGo keeps software projects moving. Gemini suggests the next safe step, scoped coding agents do approved work, local validation checks the result, and a human decides when it is done.

Comment

Inspiration

This work-in-progress project started with a question I kept asking myself as a developer: how can I manage my time better when planning, coding, testing, and keeping track of what has already been done all compete for attention?

AI tools can help with individual parts of the work, but they usually solve only the middle of the job. Someone still has to decide which task is ready, check its dependencies, give the agent enough context, verify what actually changed, run the real tests, and update the project plan afterward. On a larger project, those handoffs can become more work than the code change itself.

That led me to AIGo: a place where useful AI capabilities and the necessary management layer can work together. It keeps a record of what was requested, what an agent actually did, what was validated, and what still needs a human decision. The competition MVP is focused on software projects, but the underlying idea could also help individuals and teams coordinate other kinds of multi-step work without losing track of progress.

I did not want to build another chat window that says a task is complete because the model said so. I wanted a project manager that could use agents while keeping authority, evidence, and the final decision clear.

What it does

AIGo organizes a software project around a five-state workflow:

Backlog → Ready → Agent Working → Review → Done

The Board shows the current tasks, dependencies, priorities, and project progress. The Advisor combines deterministic project checks with Gemini suggestions. For example, AIGo can identify a milestone risk from the deadline and unfinished tasks, then ask Gemini for a structured planning proposal. The proposal is still only advice. A human reviews it, and AIGo checks the latest project revision and workflow rules again before applying anything.

Daily Watch handles small recurring research jobs such as following relevant laws, events, news, or milestone changes. External findings use Gemini with Google Search grounding and retain their source citations. Each watch is eligible at most once per UTC day, has an end date capped at one month, and never changes a task automatically.

For coding work, Antigravity 2.0 connects to AIGo through MCP. Its token is limited to one project and a selected set of coordination tools. It can inspect eligible work and request a claim, but it does not receive repository paths, arbitrary shell access, or permission to approve its own result.

Repository work happens through a separate Windows Local Bridge. The bridge uses its own device credential and an explicitly approved repository and validation profile. It starts the configured Codex CLI runtime locally, observes which files changed, and runs host-owned validation. Only a sanitized result is sent back to the cloud. A validated run may move the task to Review; only a human can move it to Done.

How I built it

I started with the workflow rules rather than the UI or the prompt. Task transitions, dependencies, stale-revision checks, and the rule that only a human owns Done live in the .NET domain and application layers.

The web interface is built with React 19, TypeScript, and Vite. The backend is an ASP.NET Core application on .NET 10, deployed to Google Cloud Run. Cloud SQL for PostgreSQL owns remote accounts, projects, tasks, proposals, and agent-run state. Firebase Email/Password Authentication provides browser identity.

Gemini is accessed through the official Google GenAI .NET SDK. AIGo sends a limited project snapshot and expects structured output that can be checked before use. Gemini does not receive filesystem or shell access and cannot write directly to the database.

The MCP endpoint is part of the ASP.NET backend. Antigravity is the external MCP client used in the demo. The Local Bridge and its SQLite database stay on the trusted computer, where repository bindings, validation profiles, and detailed local evidence belong.

For Daily Watch, Cloud Scheduler starts a short private Cloud Run Job. The job processes a limited daily batch and exits instead of keeping an agent running continuously.

Data sources

AIGo's main source of truth is the project itself: tasks, dependencies, milestones, revisions, agent runs, and human decisions stored in PostgreSQL.

External Daily Watch checks use Google Search grounding through Gemini. AIGo accepts public links only from the SDK's grounding metadata and stores them as citations. Milestone-risk checks are calculated directly from project data and do not require a web search.

Challenges

The hardest part was not making a Gemini request. It was deciding what the model and each agent should be allowed to do.

Planning suggestions can become stale while someone else edits a project, so every approved action has to be checked against the latest project revision. Coding introduces a different problem: a successful agent process does not prove that the right files changed or that the application still works. I had to keep the coding-process result, changed-path policy, build and test results, remote workflow state, and human acceptance as separate pieces of evidence.

The cloud/local boundary was another challenge. The web application needs enough information to show useful progress, but local paths, commands, raw logs, source code, and credentials should not be copied to the server. The Local Bridge therefore keeps detailed evidence locally and sends only a normalized summary.

I also wanted Daily Watch to feel like a long-running agent without creating an open-ended token bill. The final design uses one short scheduled batch, a daily eligibility rule, a small project limit, and a mandatory expiry date.

Accomplishments that I am proud of

I am most proud that AIGo demonstrates a complete loop rather than a collection of disconnected AI features:

  1. It finds a useful piece of work.
  2. Gemini can suggest a structured plan change.
  3. A human approves the change.
  4. A scoped MCP client claims an eligible task.
  5. A separately authenticated local process performs and validates the work.
  6. The result stops at Review.
  7. A human decides whether it is actually Done.

The product also remains useful if Gemini is temporarily unavailable. The board and deterministic project analysis still work, and tasks can still be created manually.

What I learned

The most useful lesson was that typed model output is much easier to trust than free-form instructions. It can be validated, rejected, retried, and recorded without pretending that model text is authority.

I also learned that agent success and project completion are different events. A tool reporting success is useful progress information, but it is not a substitute for independent validation or human acceptance.

Finally, a long-running agent does not have to be an always-running process. For this project, a small scheduled job with explicit limits and an expiry date is simpler, cheaper, and easier to explain.

What's next

Next I want to make Local Bridge setup easier with a clearer connection test and better recovery guidance. I also want reusable validation-profile templates, stronger run observability, and a better way to compare grounded Daily Watch findings over time.

I also want to add epic tracking so larger project goals can be broken into milestones and related tasks, with progress visible at every level. Another important next step is multi-user collaboration: several people should be able to work on the same project, while task ownership, reviews, approvals, and the history of each decision remain clear.

The main rule will stay the same: agents may help plan and execute work, but they should not silently create their own authority or declare themselves finished.

Development disclosure

AIGo was created during the hackathon submission period, beginning August 17, 2026. OpenAI Codex was used as a development assistant. AIGo’s qualifying Google agent framework is the official Google GenAI SDK. The demo uses Antigravity 2.0 as an MCP client.

Built With

Share this project:

Updates

Submission history