From broad goals to the first agent shipped in five minutes.
My first PRD! I wrote this in the winter of 2026, and the landscape of agentic AI has changed dramatically since, but it was a fantastic first exposure to the product space.
Author
Noah Proctor
Product
Aden Hive (YC W25)
Type
Feature PRD
Context
Aden's Hive is an open-source, YC-backed framework for building autonomous AI agent "swarms", one example of a bee motif we see throughout the product. Instead of manually coding workflows, developers describe a goal in natural language. Hive generates the entire agent graph, including connection logic, using a "queen" coding agent that oversees specialized worker agents. Self-healing, real-time observability via WebSocket streaming, and model-agnostic LLM support (OpenAI, Anthropic, Gemini) support it's position as an efficient, effective, and customized means of developing an agent.
Technical differentiation is tangible. Goal-driven generation means no more hardcoded DAGs (directed acyclic graph). The adaptive loop, where the queen captures failures and evolves the agent graph automatically, is a step beyond static orchestration frameworks like LangGraph or CrewAI.
The problem: Hive's strengths are invisible until you've already committed significant setup time. The current onboarding asks developers to understand queen-worker architecture, configure their LLM, and define a goal---all before seeing a single agent run. This process feels intuitive for developers, but not for a segment of users without technical experience, which generates mountains of friction in their company.
Problem Statement
New users land in one of two buckets: curious developers who found Hive on GitHub or Twitter and technical buyers evaluating it for production use. Both share a common inflection point; they churn before reaching the "aha" moment.
The "aha" moment for Hive is watching your natural-language goal turn into a running multi-agent system in real time. Although this moment is super valuable, the time-to-value today is too long: users must clone the repo, configure environment variables, understand the architecture, and write their first goal entirely cold, without guidance or templates.
| User Type | What They Need | What They Get Today |
|---|---|---|
| Curious dev Found via GitHub/X, exploring |
Immediate proof the magic works. A 90-second "wow" before any commitment. | A README with architecture diagrams, setup instructions, and no live demo. |
| Technical buyer Evaluating for production |
Trust signals: security, reliability, team credibility, enterprise use cases. | Open-source codebase with no case studies, no SOC 2 mention, no testimonials. |
| Builder Ready to ship something real |
Scaffolding for common patterns: data pipelines, research agents, browser automations. | Blank slate. Must invent their goal from scratch with no reference implementations. |
Goals & Success Metrics
This PRD targets a single North Star: get users to their first agent run faster, and make that run memorable enough that they return.
| Metric | Baseline (Est.) | Target (90 days) |
|---|---|---|
| Time-to-first-agent-run (new users) | > 30 min | < 5 min |
| D7 retention (users who ran ≥1 agent) | Est. < 20% | > 40% |
| GitHub stars / Discord joins from sandbox | – | Trackable attribution via UTM |
| Template gallery usage rate | N/A | > 60% of new users start from a template |
Note: Baseline metrics are estimates based on open-source project benchmarks and direct product observation. Instrumentation should be implemented alongside these features.
Proposed Solution
1. Live Sandbox — No Setup Required
A browser-based playground where visitors describe a goal and watch Hive generate and execute an agent swarm in real time, without installing anything. Pre-warmed execution environments, capped compute (3-minute timeout), and read-only tool access (no credential requirements) de-risk this for the team.
The sandbox is the pitch deck. Investors, developers, and enterprise evaluators should all be able to see the product work before reading a single line of docs.
2. Goal-First Setup Wizard
Replace the current cold-start with a 3-step conversational wizard. Rather than asking users to configure infrastructure before they know if Hive will solve their problem, lead with the goal.
| Step | Prompt | Output |
|---|---|---|
| 1 | "What do you want to automate?" | Hive queen previews the agent architecture it will generate, in plain English. |
| 2 | "Which tools does your agent need access to?" (checkboxes: CRM, web browsing, database, email, etc.) | Auto-generates the MCP tool configuration. User only adds credentials for selected tools. |
| 3 | "Which model should power your queen?" (dropdown with recommendations per use case) | Final config file ready. One-command launch. |
3. Template Gallery
A curated library of 10–15 production-ready agent templates organized by vertical: Sales (lead research + CRM update), Ops (invoice processing), Engineering (codebase Q&A agent), Research (competitive intelligence), and Growth (content pipeline). Each template ships with a goal description, sample tool config, and a short video of it running.
Templates lower the creative burden for new users and serve as implicit proof-of-concept for enterprise buyers evaluating specific use cases. They should be community-extensible (GitHub PRs) to scale the library without internal maintenance cost.
User Stories
Edge Cases & Open Questions
| Question | Current Lean |
|---|---|
| Should the sandbox require signup to run? | No. Friction kills the demo, so capture email post-run with a "save your result" hook. |
| How do we prevent sandbox abuse (infinite loops, prompt injection)? | Hard timeout at 3 minutes, no external HTTP calls in sandbox mode, and the output sanitized. |
| Should templates be opinionated (single LLM) or model-agnostic? | Offer a default (GPT-4o or Claude Sonnet) with a "change model" override, which educes decision fatigue. |
| Does a wizard conflict with the "describe your goal naturally" promise? | The wizard is scaffolding for configuration, not for the goal itself. The goal field remains freeform. |
| How do we handle a goal the queen can't confidently architect? | Queen responds with a clarifying question (same UX as goal refinement loop). Never silently fails. |
Prioritization
| Feature | Priority | Rationale |
|---|---|---|
| Template Gallery | P0 | Lowest engineering lift. Immediate impact on builder retention. Ships in one sprint. |
| Goal-First Setup Wizard | P0 | Directly removes the highest-friction moment in the current funnel. No new infrastructure required. |
| Live Sandbox (no-signup) | P1 | High impact on top-of-funnel conversion. Requires sandboxed execution environment; this is non-trivial infra work. |
| Trust signals page (SOC 2 roadmap, case studies, testimonials) | P1 | Unblocks enterprise sales conversations. Content-heavy, low eng cost. Needs 2–3 design hours. |
| Run history & saved goals UI | P2 | Useful for retention but depends on auth system. Lower urgency than acquisition improvements. |
Out of Scope
This PRD was written by Noah Proctor as part of an independent product review of Aden Hive. It is not an official Aden document. Observations are based on public product research, the open-source repository, and direct use of the framework. Originally published as a thread on X — @noah_proct.