Dense Slides: Automating Large Scale Refactors with Parallel Agents - Robert Brennan, OpenHands

Source Video

Automating Large Scale Refactors with Parallel Agents - Robert Brennan, OpenHands

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: agent_vision.

Slide text:

About Us

Robert Brennan

CEO and Co-founder, OpenHands

OpenHands

An MIT-licensed coding agent

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense small bullet text and multiple year sections are better handled by OCR than direct vision transcription.

Slide text:

? @

6 uthopr Autamtubg khe Aiecten tn Agtrumlt bo (-- - 1

A Brief History of LLM Coding

2022: Context-unaware code snippets

"Write python for bubble sort"

2023: Context-aware code generation

"Write' a suite of unit tests for this function"

2024: Single agents for atomic coding tasks..

Implement'the GET /products endpoint and use

curl to validate that it works".

2025: Parallel agents for large-scale work:

"Migrate our app from Redux to React Query"

2025-11-22114:35:43

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.96
Text source: agent_vision.

Slide text:

Evolution of AI-Driven Development

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: agent_vision.

Slide text:

Why do we need orchestration?

Why can't today's agents one-shot large tasks?

Agent Problems

Limited context window

Laziness

Lack of domain knowledge

Compounding errors

Human Problems

Can't convey intuition

Difficulty decomposing tasks

Need intermediate reviews/check-ins

Ambiguous definition of done

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.97
Text source: agent_vision.

Slide text:

Agentic Engineers

Every dev will use agents for day-to-day work.

Only ~5% will become Agent Orchestrators.

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: agent_vision.

Slide text:

Human in the Loop: Single Agent

Human prompts agent

Agent does some work

Human checks output

Recreated text/layout view: open HTML recreation
AI slide classifier: demo_video confidence 0.94
Text source: agent_vision.

Slide text:

Showcase: Automating Massive Refactors

Recreated text/layout view: open HTML recreation
AI slide classifier: demo_video confidence 0.93
Text source: agent_vision.

Slide text:

Showcase: Automating Massive Refactors

Recreated text/layout view: open HTML recreation
AI slide classifier: product confidence 0.98
Text source: advanced OCR rapidocr-live/bright-screen/contrast.
OCR decision: ready — Dense UI screenshot with many small labels and metrics; OCR is better than direct transcription.

Slide text:

csmith49/OpenHands

Dependency. Obce05Details Edkt Metadats

02a7da Status: NEW Vecify Flx

VCode Metrics (4 files)

ffd83d -0.2+ 66.0 CyciomaticCorpiety Maintarabity inder 59.6

Obce05 40.0.5t 593 Hangtood Ertort Hahtnad Vohune 13.8 Halstead Diffcufty Estimuated Blugs

b794a3 8166 0.20

n.0-0

Dependencies(o)

Dependenb (5)

Nodes(4)

openhands/apo_server/units/sg_siis.py 小

F t Ne

openhandsopp_serveriutils/async_remote_workspace.py

2025-11-22 14:54:08

Recreated text/layout view: open HTML recreation
AI slide classifier: title_card confidence 0.99
Text source: agent_vision.

Slide text:

Task Decomposition

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: agent_vision.

Slide text:

Task Decomposition

Break your end-goal down into tasks that:

- Are one(-ish) shottable

- Fit in a single commit

- Can be executed in parallel

- Can be verified by a human as correct or incorrect

- Clear dependencies/ordering between tasks

Recreated text/layout view: open HTML recreation
AI slide classifier: code confidence 0.98
Text source: none.
OCR decision: ready — Dense code screenshot with small source text; OCR is the right extraction path.

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.97
Text source: agent_vision.
OCR decision: ready — Product/docs screenshot with dense small UI text and code snippets.

Slide text:

Getting Started

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense slide text with long prompt, URLs, and configuration details that are better OCRed than transcribed by eye.

Slide text:

?

0. Wertahopr Autorstrg Lmhe Alectert wit Agtrls 1l t ho 心 Domg + D

+心:t4心:t4h +.4

Vibecoding CVE Remediation: LLM Connection

Prompt 1:

Look at the example code at https://docs.openhands.dev/sdk/auides/hello-world and the SDK docs at: https://docs.openhands.dev/sdk/api-reference/openhands.sdk.llm

a pydantic SecretStr). Set the model name to: connection to the OpenHands LLM using the LLM API KEY env var (be sure to use. openhands/claude-sonnet-4-20250514 Write a small python script called cve_solver.py that for now just tests the 中

Note to human: replace model name if using anthropic/openai

Ct 2025-11-22115:19:54

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense instructional slide text with bullets, a repository URL, and environment-variable references.

Slide text:

?

nm

D Wariahopr Autoretrg tthe Aecten tthAgtrba M t bo 6o 00○0：θ■0:

Vibecoding CVE Remediation: Detect CVEs

Prompt 2:

Look at the example code at https://docs.openhands.dev/sdk/guides/hello-world

Modify cve_solver.py so that it.

Takes in a GitHub repository as a command line argument Creates a RemoteWorkspace connecting to 1ocalhost:8000 Clones the given repository to that workspace, using the GITHUB_ TOKEN env

Tells the agent to use trivy to'scan the repository for CVEs Creates an Agent in that workspace var.

CVEs Human: you can test the script with github.com/rbren/polaris - It should find 3

20251.1-22115:30:40

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: none.
OCR decision: ready — Slide has mixed prose and a code screenshot; OCR is the safer way to capture the exact agent setup and command text.

Hidden Non-Slide Evidence

`slide-016.jpg` — demo_video confidence 0.89; Mixed screen capture with a cropped slide on the left and terminal/demo output on the right; not a clean readable presentation slide.

Classification audit: raw/sources/slide-ai-classification/dense/rcsliSIy_YU/audit.json