BountyOps Autopilot

Inspiration

Small builder teams do not usually fail because they lack ideas. They lose time because opportunity discovery, eligibility checks, deadlines, build planning, verification, and submission work live in different places. BountyOps Autopilot turns that uncertainty into an executable workflow inside Slack.

What it does

BountyOps accepts an ambiguous goal such as "find the strongest cash-backed agent opportunity we can finish this week." It searches current opportunities, ranks them by deadline, reward, fit, and risk, and produces a do-now / backup / skip recommendation.

After a team chooses an opportunity, BountyOps creates an end-to-end plan with five phases: qualify, plan, build, verify, and publish. It identifies required assets, orders work by dependency and deadline risk, and keeps consequential actions behind a human approval gate.

How we built it

• Slack and Bolt for JavaScript provide the conversational work surface.
• The OpenAI Agents SDK orchestrates the agent and local tools through an OpenAI-compatible API.
• The Qwen Cloud provider adapter selects Qwen models through environment configuration.
• scan_cash_hackathons discovers and ranks real opportunities.
• prepare_autopilot_execution_plan generates build, evidence, verification, and approval phases.
• The repository includes an Alibaba Cloud deployment path using Container Registry, ECS, and the ECS RunCommand API.

How Qwen Cloud fits

Qwen Cloud is designed as the reasoning layer rather than a text-only wrapper. It interprets incomplete goals, selects tools, combines structured tool results, explains tradeoffs, and produces a concrete next action. The agent must handle changing deadlines, inconsistent prize formats, eligibility constraints, and public-action risk without forcing the user through a fixed questionnaire.

At submission preparation time, the Qwen Cloud account's hackathon model-access request was still pending, so the public repository and demo distinguish verified BountyOps behavior from the provider deployment path instead of presenting an unverified cloud call as completed.

Challenges

The hardest part was separating useful autonomy from unsafe automation. An agent that can discover and plan work should not silently spend money, sign a wallet message, publish a repository, or submit an entry. We encoded those boundaries in the system prompt and in the execution plan itself.

Another challenge was making the workflow reproducible. We added deterministic ranking logic, tests, a provider verification command, secret scanning, architecture documentation, and an Alibaba Cloud deployment script.

Accomplishments

• A Slack-native agent that converts open-ended requests into ranked execution decisions.
• Real tool invocation instead of a scripted chat interface.
• Explicit human approval checkpoints for consequential actions.
• Automated tests and submission-readiness checks.
• A public Apache-2.0 repository with documented Qwen and Alibaba Cloud integration paths.

What we learned

The best agent workflow is not the one with the most autonomy. It is the one that knows which decisions can be automated, which facts must be verified, and exactly where a human must take responsibility.

What's next

We plan to complete the live Qwen Cloud verification when model access is approved, deploy the container to ECS, add persistent execution state, monitor deadline changes, and build team-level Slack canvases that track evidence and approvals across an entire submission cycle.

Built With

• alibaba-cloud
• bolt
• javascript
• node.js
• openai-agents-sdk
• qwen-cloud
• slack