Last week, we ran AI AgentHack, a hackathon where more than 3,000 developers built creative agentic projects on Portia. Picking winners wasn’t easy, but one project stood out: Team Dark Mode’s DCRCA Agent (Disaster Chaos Response Coordination AI).
The DCRCA Agent helps emergency teams cut through the noise. It scans live news and social feeds, pulls out the key details, and maps emergencies by priority so responders know exactly where to act first.
We were extremely impressed to see this cool Portia use case and, more importantly, an application with great potential for societal impact!
Below is a deep dive into how the team built this using Portia.
The DCRCA Agent is wired together with PlanBuilderV2, where the workflow is laid out step by step: pulling raw data from news feeds, parsing and prioritizing it with LLM steps, routing through a human approval checkpoint, and finally dispatching updates over email and Slack. Each stage has clear inputs and outputs, making the whole flow transparent.
A key design choice was the separation between reasoning and tool calls. Reasoning tasks (like structuring raw data or scoring emergencies) live inside .llm_step()
, while external services such as Google Search and Gmail are called through .invoke_tool_step()
. This separation keeps debugging and maintenance straightforward.
They also used custom Python functions for oversight with .function_step()
. These functions handled approval checks and message formatting, showing how Portia makes human-in-the-loop workflows natural instead of forcing full automation.
Finally, because every step exposes structured outputs at runtime, the agent can surface both intermediate results (like “Slack message sent ✅”) and the overall summary of actions — giving the team visibility into exactly what happened.
We’re thankful to Team Dark Mode and all other hackathon participants for helping us prove that Portia isn’t just for tinkering—it can drive real, high‑stakes workflows. By combining off‑the‑shelf tools, LLM reasoning and human oversight in a single plan, they built something useful and understandable.
It’s exciting to imagine what other novel agentic ideas the community will bring to life next!
If you want to try building your own agentic workflow, check out our GitHub.
Top comments (0)