By Adam D’Angelo, Vice President of Technology Solutions, Acuity

TSG researched a wide set of tools across this emerging landscape, including:

Devin, RooCode, Cline, Cursor, Windsurf, Super‑Claude, Claude Code, Zed, 

Copilot Agent, Amazon Kiro, Gemini CLI, and Vercel v0.

While each has a unique approach, one theme was consistent: these AI tools are no longer just copilots. They’re teammates.

The Shift: From Code Suggestions to AI Agents

Legacy autocomplete tools focused on the next token or line of code. Today’s top platforms (especially Devin) take it many steps further. Devin acts like a junior developer: it reads a user story, generates specs, builds a plan, executes code changes, runs tests, and opens pull requests autonomously.

Devin topped our evaluation for its end-to-end autonomy, context memory, DevOps integration, and fine-grained approval flows. It’s a powerful demonstration of how far AI agents have come. From helpful typists, they now work as hands-on executors.

Planning First: Spec Generation and Task Breakdown

One of the most useful evolutions is the “plan-first” workflow. Tools like Devin, Amazon Kiro, and Claude Code emphasize:

• Turning plain-language requirements into structured specs
• Breaking down specs into executable tasks
• Generating designs, data models, and architecture diagrams
• Surfacing missing or ambiguous details early in the process

This leads to faster alignment with stakeholders and fewer surprises during implementation.

Case Study: A Mortgage Loan App, Built by AI

To put these capabilities to the test, we turned to Subject Matter Expert Chris Sahno, a member of Acuity’s Technology Solutions Group. Chris built an entire mortgage loan application using only AI-generated code. He started with a Markdown file containing simple, human-readable requirements describing workflows for customers and loan officers, data validation, and security needs.

The AI (in this case, Kiro) translated those requirements into a full technical spec, then:

• Refactored the codebase from vanilla HTML/CSS into Material UI
• Implemented role-based access controls and real-time status updates
• Handled secure data processing with audit logging
• Generated tests and summaries for every task it executed

Chris even introduced bugs and new features midstream to test the AI’s responsiveness. The tool adapted in real time, demonstrating how AI development agents can handle live iteration cycles.

Autonomy with Oversight: Multiple Execution Modes

These tools offer varying levels of autonomy, from manual, step-by-step execution to auto-mode, where the AI works in the background and submits code for review.

Devin again leads in flexibility, allowing users to supervise as needed while the AI:

• Runs tests
• Executes shell commands
• Updates multiple files across the repo
• Opens pull requests with unified diffs and rollback options

This combination of productivity and control is key to trust and adoption.

Comparing the Field

Here’s how the tools stack up based on our evaluation:

What’s Next for Us

We’re now focused on how to incorporate these tools into our delivery workflows. Next steps include:

• Defining AI-assisted SDLC patterns for legacy modernization
• Hosting internal events on prompt engineering and tooling strategies
• Evaluating team size vs. automation trade-offs
• Developing internal wrappers or integration layers to enhance consistency and quality

We believe this wave of AI tooling will change not just how we write software but who writes it and how teams organize around it.

Learn more about Acuity’s expanding AI expertise in our recent blog Unlocking Innovation: Lessons from the AWS LLM League 2025.