How to Build a Multi-Agent Review Swarm: Summary & Key Takeaways | VibeDigest

How to Build a Multi-Agent Review Swarm

youtube.com12 min

Summary

A Notion engineer demonstrates an AI-native development workflow that turns spoken intent into detailed tasks, cloud-executed code, automated reviews, passing CI, and pull requests. The main productivity gain comes from rigorous planning, persistent shared context, and self-correcting review loops—not from expecting flawless one-shot generation.

A Notion engineer demonstrates an AI-native development workflow that turns spoken intent into d…Read more

Key ideas

Speech captures richer implementation context
The engineer starts nearly every task by speaking rather than typing. Free-form narration lets him include small details, corrections, uncertainty, and changing preferences without editing them into a polished prompt. He believes the model can tolerate verbal noise and may even benefit from seeing how the request evolves. The resulting prompt is therefore closer to a complete stream of engineering intent than a terse written command.
Why it matters: Voice input can reduce the effort required to specify a task while improving the contextual richness available to the agent.
Supporting evidence
I'm like a big believer in if I'm like speaking, there's all these like little bits of like nuance and detail uh that I'm going to provide versus me just like typing in the keyboard.
The engineer starts nearly every task by speaking rather than typing. Free-form narration lets h…Show more
The engineer starts nearly every task by speaking rather than typing. Free-form narration lets h…Show moreShow less
The engineer starts nearly every task by speaking rather than typing. Free-form narration lets him include small details, corrections, uncertainty, and changing preferences without editing them into a polished prompt. He believes the model can tolerate verbal noise and may even benefit from seeing how the request evolves. The resulting prompt is therefore closer to a complete stream of engineering intent than a terse written command.
Why it matters: Voice input can reduce the effort required to specify a task while improving the contextual richness available to the agent.
Supporting evidence
I'm like a big believer in if I'm like speaking, there's all these like little bits of like nuance and detail uh that I'm going to provide versus me just like typing in the keyboard.
Knowledge and compute are joined in one workspace
Notion AI can use Boxy, an internal system of prebuilt remote coding machines, instead of beginn…Show moreShow less

Read the full digest

Overview

The interview shows what software development looks like when an AI agent is embedded in both a company’s knowledge base and its cloud development environment. The engineer begins projects by speaking freely, capturing more nuance, constraints, and corrections than he would typically type. Notion AI then connects to preconfigured remote coding machines called Boxy, explores the real codebase, and converts the discussion into a detailed task containing requirements, code pointers, and verification steps. A separate execution thread implements the task, invokes a custom review-swarm skill, opens a pull request, monitors continuous integration, and fixes routine failures without requiring the engineer’s attention. Notion simultaneously serves as the canonical source for specifications and as a living journal that records discoveries and problems during implementation. The workflow reduces application switching and eliminates local-environment maintenance, while allowing long-running work to continue in the cloud during meetings. Its most important lesson is that successful agentic coding depends on investing human effort in goals, environment, constraints, and validation. The engineer’s custom review system reinforces this principle by dividing changes into technical domains, evaluating both correctness and maintainability with multiple models, consolidating findings, and iterating until the change is ready for human review.

Sections

Higher-Order Insights

Broader implications of embedding coding agents inside a shared knowledge and execution environment.

The workflow treats context management as the central productivity constraint. Its advantage comes from co-locating specifications, organizational knowledge, execution infrastructure, progress notes, and delivery status so neither the engineer nor the agent must repeatedly reconstruct state.
Agent autonomy here is bounded rather than absolute: the human defines outcomes and constraints, while the agent owns execution, review, CI monitoring, and routine correction. This suggests a practical division of labor in which judgment remains human-led and repetitive feedback loops become machine-led.
The detailed task page functions like an executable contract between planning and implementation. Because it includes requirements, code pointers, and verification criteria, it can coordinate multiple agent stages while remaining legible to humans.
Pinned chat threads form an informal personal work queue, revealing that agent-native workflows may replace some conventional project-management behavior without replacing the company’s canonical task system.

Practical Next Steps

Concrete ways to adopt or evaluate the demonstrated workflow.

Pilot voice-first task intake for one substantial engineering change, ensuring the narration covers the desired outcome, environment, affected behavior, constraints, and acceptance criteria.
Create a persistent planning artifact for every delegated implementation that includes code pointers, explicit requirements, and verification steps before allowing the coding agent to begin.
Prepare a reproducible remote development environment with repository access and required tooling so agents do not repeatedly begin from an empty sandbox.
Automate the post-generation loop so the agent runs reviews, opens a pull request, monitors CI, fixes routine failures, and records unresolved issues for human attention.
Design review roles around distinct objectives such as correctness and maintainability, then compare their findings with the results of a single generic reviewer before expanding the system.
Require the agent to maintain a work journal containing discoveries, problems, decisions, and verification outcomes in the same workspace as the specification.

Technical Implementation Details

Specific systems, components, and orchestration patterns described in the interview.

Notion AI can provision a computer for complex work and currently uses Vercel sandboxes for general-purpose command execution and artifact creation.
Boxy is the internal nickname for prebuilt cloud coding VMs that provide remote compute on demand and access to the codebase.
The planning output includes code pointers, migration requirements, and self-verification steps, and can be added directly to the company’s canonical task system and milestone.
The execution agent is instructed to satisfy the task requirements, invoke the review-swarm skill, open a pull request, monitor CI, and repair type errors or failing tests.
The review swarm partitions the change set by domain, assigns correctness and maintainability reviewers within each domain, runs each slice through GPT and Opus, and uses a top-level agent to consolidate actionable findings.
The review loop repeats until the consolidated findings are addressed, while the agent writes implementation notes back into Notion as a living work journal.

In Brief A Notion engineer demonstrates an AI native development workflow that turns spoken inte…Expand full summary

Overview

Sections

Higher-Order Insights

Broader implications of embedding coding agents inside a shared knowledge and execution environment.

The workflow treats context management as the central productivity constraint. Its advantage comes from co-locating specifications, organizational knowledge, execution infrastructure, progress notes, and delivery status so neither the engineer nor the agent must repeatedly reconstruct state.
Agent autonomy here is bounded rather than absolute: the human defines outcomes and constraints, while the agent owns execution, review, CI monitoring, and routine correction. This suggests a practical division of labor in which judgment remains human-led and repetitive feedback loops become machine-led.
The detailed task page functions like an executable contract between planning and implementation. Because it includes requirements, code pointers, and verification criteria, it can coordinate multiple agent stages while remaining legible to humans.
Pinned chat threads form an informal personal work queue, revealing that agent-native workflows may replace some conventional project-management behavior without replacing the company’s canonical task system.

Practical Next Steps

Concrete ways to adopt or evaluate the demonstrated workflow.

Pilot voice-first task intake for one substantial engineering change, ensuring the narration covers the desired outcome, environment, affected behavior, constraints, and acceptance criteria.
Create a persistent planning artifact for every delegated implementation that includes code pointers, explicit requirements, and verification steps before allowing the coding agent to begin.
Prepare a reproducible remote development environment with repository access and required tooling so agents do not repeatedly begin from an empty sandbox.
Automate the post-generation loop so the agent runs reviews, opens a pull request, monitors CI, fixes routine failures, and records unresolved issues for human attention.
Design review roles around distinct objectives such as correctness and maintainability, then compare their findings with the results of a single generic reviewer before expanding the system.
Require the agent to maintain a work journal containing discoveries, problems, decisions, and verification outcomes in the same workspace as the specification.

Technical Implementation Details

Specific systems, components, and orchestration patterns described in the interview.

Notion AI can provision a computer for complex work and currently uses Vercel sandboxes for general-purpose command execution and artifact creation.
Boxy is the internal nickname for prebuilt cloud coding VMs that provide remote compute on demand and access to the codebase.
The planning output includes code pointers, migration requirements, and self-verification steps, and can be added directly to the company’s canonical task system and milestone.
The execution agent is instructed to satisfy the task requirements, invoke the review-swarm skill, open a pull request, monitor CI, and repair type errors or failing tests.
The review swarm partitions the change set by domain, assigns correctness and maintainability reviewers within each domain, runs each slice through GPT and Opus, and uses a top-level agent to consolidate actionable findings.
The review loop repeats until the consolidated findings are addressed, while the agent writes implementation notes back into Notion as a living work journal.

How to Build a Multi-Agent Review Swarm

Summary

Key ideas

Speech captures richer implementation context

Knowledge and compute are joined in one workspace

Planning is separated from execution

Demonstrated Agentic Development Loop

Ask about this source

Read the full digest

Overview

Sections

Higher-Order Insights

Practical Next Steps

Technical Implementation Details

Overview

Sections

Higher-Order Insights

Practical Next Steps

Technical Implementation Details