WorkPilot

Give AI the job, not the prompt.

Comment

WorkPilot — Give AI the Job, Not the Prompt

WorkPilot is a professional AI agent that turns a high-level business objective into an executable workflow instead of requiring the user to engineer a complex prompt.

The Problem

Most AI tools still make the human responsible for breaking a complicated job into prompts, deciding what information to gather, checking whether sources are trustworthy, resolving conflicting evidence, and validating the final answer.

That approach does not scale well for professional work.

A real research task is a workflow:

Goal → Scope → Research → Evidence → Analysis → Human Review → Synthesis → Quality Check

WorkPilot is designed around that workflow.

What We Built

WorkPilot accepts a natural-language job such as:

Research Stripe, Adyen, and Paddle and produce a competitive-intelligence brief comparing their products, pricing, target customers, features, positioning, strengths, weaknesses, and the best fit for different business situations.

Instead of asking the user to construct the prompts and sequence manually, WorkPilot creates and executes the workflow.

The agent:

  1. Defines the scope and research criteria.
  2. Plans executable tasks.
  3. Performs live web research.
  4. Prefers primary sources when gathering evidence.
  5. Tracks evidence and uncertainty.
  6. Analyses findings and trade-offs.
  7. Detects situations where human judgement is valuable.
  8. Stops at a human approval gate before final synthesis.
  9. Continues only after the human approves.
  10. Runs a final quality check.
  11. Produces a structured, decision-ready report.

Why Human Approval Matters

WorkPilot is intentionally not a fully autonomous black box.

When the agent reaches a point where evidence or uncertainty could affect the final recommendation, it requests human approval.

The human can:

  • Approve and continue
  • Reject and stop

This creates a deliberate control point between machine reasoning and consequential output.

The goal is not simply to make AI autonomous.

The goal is to make AI usefully autonomous while keeping humans in control.

AWS Strands Agents

WorkPilot is built around the AWS Strands Agents SDK as the agent orchestration layer.

Strands is used to structure the agent workflow, execute phase-aware tasks, and coordinate the progression from planning through research, analysis, human approval, synthesis, and quality checking.

The application also uses AWS Bedrock Mantle for model access.

Architecture

The system is intentionally lightweight:

  • TypeScript / Node.js application
  • AWS Strands Agents
  • AWS Bedrock Mantle
  • Live research/search and fetch tools
  • Evidence and research state
  • Task-based workflow orchestration
  • Human approval event/gate
  • Quality validation
  • Native HTTP demonstration interface

The architecture is designed so that the same workflow engine can support many professional jobs beyond competitive intelligence.

What We Learned

The biggest lesson was that reliable agentic systems require more than a good model response.

The workflow itself matters.

We learned to separate:

  • planning from execution,
  • research from analysis,
  • evidence from conclusions,
  • machine execution from human approval,
  • and generation from quality validation.

We also learned that phase classification is important. A research task should not accidentally be treated as a synthesis task, and a quality task should not be confused with research.

Challenges

The main technical challenge was coordinating multiple stages while preserving the context produced by previous stages.

We addressed this with explicit workflow state, task statuses, phase-aware execution, evidence tracking, event generation, and a human approval state.

Another challenge was making the system demonstrate real agent behaviour rather than looking like a static mockup. The final demo therefore runs the actual workflow and exposes the approval step before the final report is produced.

Why WorkPilot

WorkPilot changes the interaction model from:

"Here is a prompt. Figure it out."

to:

"Here is the job. Execute the workflow."

That distinction is the core idea behind WorkPilot.

It is a professional AI agent designed to move from prompts toward delegated work.

Built With

Share this project:

Updates

Submission history