Inspiration
The web was built for humans to click, read, compare, and submit. AI agents are changing that—but most websites still force agents to behave like humans: guessing interfaces, interpreting visual elements, navigating menus, and scraping information from rendered pages.
We wanted to explore a different question:
What happens when a web application is designed from the beginning for humans and AI agents to operate together?
Atlas Mission Copilot grew from that idea. We envisioned a system where a human can describe a difficult real-world objective—such as improving water reliability with a limited budget—and an AI agent can work through Atlas's structured intelligence to discover opportunities, compare interventions, inspect evidence, learn from previous failures, assess risks, prepare a plan, and ask the human for approval before taking consequential action.
The inspiration is bigger than an AI assistant. It is a new interaction model:
Human + Agent + Web Application + Real-World Systems.
What it does
Atlas Mission Copilot turns a natural-language objective into a structured, auditable workflow.
A human starts with a goal such as:
“We have a limited budget. Find the highest-priority infrastructure problem, compare interventions, learn from similar failed projects, build the strongest pilot, and show me what you still don't know.”
The agent can then use Atlas's WebMCP tools to:
Discover relevant missions, projects, communities, assets, problems, and evidence.
Understand project state, infrastructure, history, and context.
Compare alternative interventions and historical outcomes.
Verify evidence, provenance, confidence, and uncertainty.
Design scenarios, intervention plans, milestones, monitoring plans, and evidence plans.
Coordinate assets, operators, partners, and field workflows.
Prepare decision briefs, pilot proposals, stakeholder briefs, and capital requests.
Measure and learn from outcomes, failures, maintenance history, and similar projects.
Crucially, Atlas separates actions into three levels:
READ — safe information retrieval.
PREPARE — analyses, scenarios, and drafts that a human can review.
EXECUTE — consequential actions that require explicit human approval.
This means the agent does not silently make decisions on behalf of the user. The human remains the decision-maker while the agent becomes the orchestrator.
Atlas also preserves the distinction between observed, reported, modeled, estimated, verified, and unknown information, so an agent can reason about uncertainty instead of presenting every result as fact.
How we built it
We built Atlas Mission Copilot on top of the existing Atlas application rather than rebuilding the product from scratch.
The central technical layer is WebMCP. Atlas exposes meaningful application capabilities through structured tools using:
document.modelContext.registerTool(...)
Rather than exposing superficial UI actions, we designed tools around real Atlas entities and workflows.
The tool surface is organized into functional groups:
Discovery: missions, projects, assets, communities, problems, and evidence.
Analysis: project details, asset status, risk, priority, history, failure patterns, and related evidence.
Decision: intervention ranking, scenario building, comparison, decision briefs, and ethics review.
Project: project drafts, intervention plans, milestones, monitoring, and evidence plans.
Field and assets: telemetry, inspections, work orders, and maintenance history.
Learning: similar projects, lessons, failures, and outcome history.
Collaboration: partner requests, capital requests, field briefs, and stakeholder briefs.
Each tool has a narrow purpose, explicit input schema, predictable output, authorization checks, read/write semantics, and error handling.
The architecture intentionally separates the agent's structured state from the human's visual experience. The agent receives structured data; the human gets clear visualizations, explanations, evidence, and approval controls.
We also designed security around WebMCP as an application boundary, including authentication, authorization, validation, rate limiting, mutation confirmation, audit trails, least privilege, and defenses against prompt injection and confused-deputy behavior.
Challenges we ran into
The hardest challenge was resisting the temptation to build “a chatbot with tools.”
It was relatively easy to connect an LLM to functions. The difficult part was deciding what those functions should actually represent.
We had to think in terms of application capabilities rather than UI buttons:
- What should an agent be allowed to discover?
- What information is safe to expose?
- How should uncertainty be represented?
- Which actions should require approval?
- How do we prevent an agent from confusing a recommendation with a verified fact?
- How do we make multi-step workflows understandable to a human watching them happen?
Another challenge was making WebMCP visible without making the interface feel like a developer console. We wanted judges and users to understand that something fundamentally different was happening without overwhelming them with implementation details.
We therefore designed an agent activity timeline showing the goal, plan, tool calls, findings, evidence, confidence, approval state, and final result.
Finally, we had to balance ambition with hackathon reality. Atlas could become an enormous platform, so we focused on one complete vertical slice rather than building dozens of disconnected features. The goal was to make one mission workflow work exceptionally well.
Accomplishments that we're proud of
We are proud that Atlas Mission Copilot demonstrates WebMCP as a fundamental product capability, not a checkbox.
The most important accomplishment is the complete loop:
Human goal → agent planning → structured tools → evidence → analysis → recommendation → human approval → action → measurement → learning.
We are also proud of the human-agent collaboration model. Atlas does not replace human judgment. Instead, it makes complex workflows dramatically easier to investigate and coordinate while keeping consequential decisions under human control.
Another major accomplishment is the emphasis on evidence and uncertainty. Atlas is designed so an agent can explicitly say that evidence is moderate, attribution is uncertain, or information is missing rather than manufacturing confidence.
We also designed a path toward physical-world coordination. Asset telemetry, maintenance history, inspections, work orders, field workflows, and real-world infrastructure can become part of the same agent-accessible system—while maintaining a hard boundary around critical hardware and requiring safe application-level actions.
Most importantly, Atlas shows a glimpse of what an agent-native web application can feel like:
Not “an AI that talks about the product.”
But a product an AI can actually use.
What we learned
We learned that the biggest opportunity with WebMCP is not simply giving agents more tools.
It is giving agents better primitives.
When capabilities are exposed as precise, structured application tools, the agent can reason over the application's actual state instead of trying to reconstruct meaning from the interface.
We also learned that human oversight is not a limitation—it is part of the interaction design. The strongest workflow is not fully autonomous. It is collaborative:
the agent explores and reasons; the human decides; the system executes safely; everyone learns from the outcome.
We learned that provenance and uncertainty need to be first-class data. A powerful agent becomes much more trustworthy when it knows not only what it found, but also where it came from, how confident the system is, and what remains unknown.
Finally, we learned that the best demonstration of agentic web technology is not a clever conversation. It is a difficult task that becomes materially easier because the agent can operate the application itself.
What's next for Atlas Mission Copilot
The next step is to expand Atlas from a compelling mission workflow into a broader agent-native operating layer for real-world projects and infrastructure.
We want to deepen the connection between project intelligence, evidence, field operations, assets, and measurable outcomes.
That means expanding agent capabilities for:
Infrastructure operations — monitoring assets, identifying anomalies, preparing inspections, and coordinating maintenance.
Field intelligence — allowing field operators to capture evidence and have agents reason over asset history, telemetry, known failure modes, and inspection requirements.
Decision intelligence — richer scenario modeling, intervention comparison, uncertainty analysis, and evidence synthesis.
Organizational coordination — helping communities, operators, funders, researchers, and implementation partners coordinate around the same mission.
Continuous learning — feeding real outcomes and failures back into the system so future recommendations improve over time.
Longer term, we see Atlas evolving toward a world where websites are not passive destinations that agents merely read. They become structured, permissioned environments that humans and agents can operate together.
The ultimate loop is:
SEE → REASON → DECIDE → ACT → MEASURE → LEARN
That is the future we want Atlas Mission Copilot to help build.
Built With
- agentic
- agents
- ai
- api
- artificial
- collaboration
- data
- decision
- development
- generative
- human-ai
- infrastructure
- intelligence
- iot
- javascript
- json
- llm
- next.js
- open
- react
- sustainability
- typescript
- visualization
- web
- webmcp
Log in or sign up for Devpost to join the conversation.