Devin 2.0 and Moore's Law for AI Agents
Scott Wu, CEO & Co-founder, Cognition AI
Scott Wu proposes that AI agent capabilities double every 70 days, creating a Moore's Law trajectory for software engineering. Learn about the dramatic evolution from tab completion (late 2023) to today's autonomous agents that can work for hours on complex tasks. Explore Devin 2.0's Deep Wiki, automated testing, and backlog processing capabilities.
"If this holds true for even just a year or two, we're talking about 16 to 64x improvement in capabilities. That's like going from barely being able to generate one line to shipping entire features."Scott Wu, Cognition AI (00:08:45)
Capability doubling
Annual improvement
Tab → Agents
Cognition valuation
Moore's Law for AI Agents
Scott Wu's framework for understanding AI agent evolution
Important Context
Note: The "Moore's Law for AI Agents" framework is Scott Wu's observation and lacks independent benchmark validation. The 70-day doubling cycle is based on Cognition's internal measurements and has not been externally verified. When citing this framework, attribute it to Scott Wu rather than presenting it as an established fact.
Time for capabilities to double
Annual improvement potential
"In the end of 2023 obviously agents were not even a concept. The biggest use case that had PMF in code was tab completion. Copilot had just come out and that was all that people thought of when they thought about AI and coding."
The state of AI coding in late 2023
Watch (00:02:30)"So we think that there is something like a Moore's Law for AI agents. The way to think about it is the capability of agents on a fixed benchmark roughly doubles about every 70 days or so."
Defining Moore's Law for AI agents
Watch (00:07:15)"If this holds true for even just a year or two, we're talking about 16 to 64x improvement in capabilities. That's like going from barely being able to generate one line to shipping entire features."
Implications of exponential growth
Watch (00:08:45)The Math Behind Moore's Law for AI
If capabilities double every 70 days:
- 1 year: 365 ÷ 70 ≈ 5.2 doubling periods → 2^5.2 ≈ 37x improvement
- 2 years: 730 ÷ 70 ≈ 10.4 doubling periods → 2^10.4 ≈ 1,400x improvement
- Practical range: 16-64x annually (conservative estimate)
18 Months of Dramatic Change
From tab completion to autonomous agents
The evolution from GitHub Copilot's tab completion (late 2023) to today's autonomous agents represents one of the fastest technological transitions in software engineering history. What seemed impossible 18 months ago is now routine.
"When you look at GPT-3, it could barely generate one line of code. GPT-3.5 got better but it was really guided. GPT-4 could finally do multi-step reasoning. And then Devin can actually work for hours on a task."
Evolution from single-line to multi-hour tasks
Watch (00:03:20)"Only 18 months ago, all the PMF in AI coding was in tab completion. That was it. Copilot had just come out and that was all that people thought of."
How quickly the landscape has changed
Watch (00:02:50)Tab Completion
GitHub Copilot launches. AI suggests single lines. No autonomy.
Chat-Based Coding
GPT-4 enables multi-step reasoning but still human-guided.
First Agents
Devin launches. End-to-end tasks with some autonomy.
Autonomous Engineers
Devin 2.0. Deep understanding, debugging, backlog processing.
Devin 2.0 Capabilities
What makes the new generation of AI agents different
Devin 2.0 represents a significant leap forward from the initial Devin release. These aren't just incremental improvements—they're fundamental capabilities that enable truly autonomous software engineering.
Deep Wiki
Maintains internal representation of entire codebase. Not just reading files one at a time, but understanding relationships and architecture.
Automated Testing
Runs tests automatically, analyzes output, debugs failures, and iterates on fixes without human intervention.
Backlog Processing
Accept entire backlog, prioritize tasks, and work through them autonomously with proper dependency handling.
Full Tool Use
Browser, terminal, editor, debugger. Uses tools like a human engineer - searching docs, running commands, testing changes.
Long-Running Tasks
Works for hours on complex tasks, maintaining context and state throughout multi-step processes.
Debug & Iterate
Core engineering skill. Doesn't just write code - analyzes failures, fixes bugs, validates solutions.
"Devin 2.0 has this thing we call Deep Wiki where it basically maintains an internal representation of the entire codebase. It's not just reading files one at a time."
Deep Wiki: Internal codebase understanding
Watch (00:12:30)"It can run tests automatically. It can look at the test output, debug failures, and iterate on fixes without human intervention."
Automated testing and debugging
Watch (00:13:15)"You can just throw your entire backlog at it. It'll figure out what to work on, prioritize, and work through tasks autonomously."
Backlog processing capabilities
Watch (00:14:20)"It has its own browser and terminal. It can actually use tools like a human engineer would - searching the web, reading docs, running commands."
Tool use capabilities
Watch (00:11:45)Technical Insights
What makes AI agents different from chatbots
Building autonomous AI agents required fundamentally new infrastructure. Traditional chat systems designed for short conversations don't work for multi-hour engineering tasks.
"The key insight is that agents aren't just about chat. They're about long-running processes that can maintain context over hours, not just minutes."
What makes agents different from chatbots
Watch (00:10:30)"We had to build entirely new infrastructure for long-running AI processes. Traditional chat infrastructure just doesn't work for multi-hour tasks."
Infrastructure challenges for long-running agents
Watch (00:11:15)"The hardest part wasn't making it code - it was making it debug. Real engineering is mostly debugging and iteration."
Debugging as the core challenge
Watch (00:15:40)Future Predictions
What's next for AI software engineering
If Moore's Law for AI agents continues, we'll see dramatic changes in how software is built. Scott Wu shares his predictions for the near future and beyond.
"In 6 months, agents will be able to take on increasingly complex tasks. We're already seeing them handle things that seemed impossible a few months ago."
Near-term predictions
Watch (00:16:20)"The real breakthrough will be when agents can collaborate - multiple agents working together on different parts of a system."
Multi-agent collaboration
Watch (00:17:10)"I think we'll see agents shift from individual productivity tools to team infrastructure. Every team will have their own agent workflows."
Agents as team infrastructure
Watch (00:18:00)Key Takeaways
Practical insights for engineers and leaders
1. Moore's Law Framework
Exponential Growth Trajectory
- •AI agent capabilities may double every 70 days (Scott Wu's observation)
- •This framework lacks independent validation but aligns with observed progress
- •If sustained: 16-64x annual improvement in capabilities
- •Going from single-line completion to entire features in 18 months
2. Rapid Evolution
18-Month Transformation
- •Late 2023: Tab completion (Copilot) was the only PMF
- •Early 2024: Chat-based coding with GPT-4
- •March 2024: Devin launches, first autonomous agent
- •Late 2024: Devin 2.0 with Deep Wiki and automated testing
3. Devin 2.0 Capabilities
Autonomous Engineering
- •Deep Wiki: Internal representation of entire codebase
- •Automated testing: Runs tests, debugs failures, iterates fixes
- •Backlog processing: Prioritizes and works through tasks
- •Full tool use: Browser, terminal, editor, debugger
4. Critical Context
Important Caveats
- •Moore's Law framework is speaker's observation, not established fact
- •No external benchmarks validate the 70-day doubling claim
- •Agents existed in late 2023 (AutoGPT, BabyAGI) but weren't production-ready
- •Independent verification needed for capability measurements
About Cognition AI
Company profile and key facts
Company Overview
Founding Team
Scott Wu
CEO & Co-founder
Steven Hao
CTO & Co-founder
Walden Yan
CPO & Co-founder
Notable Investors:
Source Video
Devin 2.0 and the Future of SWE
Scott Wu • CEO & Co-founder, Cognition AI
Research Note: All quotes in this report are timestamped and link to exact moments in the video for validation. This analysis covers Scott Wu's Moore's Law framework for AI agents, the evolution from tab completion to autonomous agents, Devin 2.0's technical capabilities, and future predictions for AI software engineering.
Key Concepts: AI agents, Moore's Law, Devin 2.0, Cognition AI, Deep Wiki, autonomous coding, SWE-bench, GitHub Copilot, tab completion, agentic AI, multi-hour tasks, automated testing, backlog processing
Related Companies
Key players in AI coding and agents
Cognition AI
Devin AI
GitHub
Copilot
Cursor
AI Code Editor
OpenAI
GPT Models