Case study
November 20, 2026

Rootstrap saves ~40% of engineering time with AI agents

Rootstrap’s engineers now hand real work, starting with code review, to Codex agents and get about 40% of their time back. Eliza got them there in a three-week sprint of 1:1 assessments and hands-on workshops.

Eliza
What Eliza delivered
Three hands-on Codex workshops and a 1:1 assessment of every engineer
Client
Rootstrap
Industry
Software
Function
Office of the CTO
Users enabled
Software engineering team
Timeline
3 weeks
~40%
Time savings
Time back across existing engineering workflows
50–60%
As adoption grows
Projected time savings as more engineers use AI agents
14
Engineers upskilled
Each assessed 1:1, then trained to close the gaps that cost them the most time
“I really appreciate both the overall feedback and the individual feedback as well from the interviews. It is thoughtful, useful, and gives us a good foundation to keep building from. I also spoke with some of the people who were interviewed, and the feedback I heard was very positive.”
Matias Mansilla
VP of Engineering @Rootstrap
The problem

Engineers used AI daily, but left most of its time savings untapped

Rootstrap’s engineers already had a solid start with AI. Most used tools like Cursor and Claude Code day to day and understood the basic workflows.

The gap was depth. Few had taken the next steps that make AI agentic, where an AI agent carries out a multi-step task on its own instead of suggesting the next line. Those steps include customizing their setup, building sub-agents and reusable skills, and connecting their tools through MCP servers so agents can work with real systems.

Code review was the biggest time sink and the clearest place to win time back. Rootstrap needed a way to measure each engineer’s AI fluency, find the gaps, and move the team toward agentic workflows that improve speed, review quality, and delivery consistency.

The solution

Three workshops to make AI agents part of everyday engineering

Eliza delivered three hands-on Codex workshops, each built around gaps found in a 1:1 assessment of every engineer. This was an enablement engagement. We trained Rootstrap’s engineers on their own workflows rather than building software for them.

Codex was the core tool, but Rootstrap’s engineers use several AI models, so the training focused on patterns that carry across tools: prompt structure, agent task design, keeping context clean, review automation, and fitting AI into existing workflows.

The three workshops ran in order, each targeting a gap from the assessment:

  1. Foundations

    Setting up AGENTS.md, the file that tells Codex how a project works, plus reusable skills, commit workflows, and Codex configuration, so agents follow each project's conventions from the first prompt.

  2. Agentic code review

    The highest-impact gap. Engineers learned to hand repetitive review work to AI agents and take back the time it consumed.

  3. Context management and MCP servers

    Giving agents the right context without overloading them, using MCP servers securely, and avoiding token bloat and oversized context files, so agents stay accurate and cost less to run.

How we did it

From assessment to roadmap in four steps

  1. Interview each engineer 1:1. Eliza spoke with 14 developers about how they use AI for coding, code review, automation, configuration, tool orchestration, and production readiness.
  2. Score AI fluency. Each engineer got an individual score, and leaders got a team summary showing exactly where training would save the most time.
  3. Run targeted workshops. The three workshops went after the biggest gaps first, starting with code review, the largest drain on engineers’ time.
  4. Leave a roadmap. Rootstrap got a clear baseline and a practical plan for moving more of its engineering work to agentic workflows.

The workshops taught patterns, not just one tool, so the time savings hold whether an engineer works in Codex, Cursor, or Claude Code.

Example use cases
Turning requirements into ready-to-build tasks
“Turn these requirements into scoped tasks with acceptance criteria. Flag open questions before coding.”
Result: Clear, reviewable work items that engineers and agents can execute without guessing.
Delegating implementation and validation
“Implement the approved plan using our project skills. Run the checks and resolve failures before handing it back.”
Before: Engineers prompt, inspect, and correct each step manually. Now: The agent carries out a bounded task and returns the change with verification evidence.
Getting an independent code review
“Review this change against the requirements. Verify suspected bugs and return actionable findings.”
Before: Engineers carry the full burden of checking AI-generated code. Now: A separate agent checks the change before human review and approval.
Turning review feedback into better habits
“Use the accepted review feedback to propose an improvement to our reusable skills.”
Result: Lessons become reviewed guidance for future work, helping prevent repeat mistakes.
The results

~40% of engineering time saved, with a clear path to 50–60%

  • Time back across the board. Rootstrap saw ~40% time savings across existing engineering workflows, with a clear path to 50–60% as adoption grows.
  • Gaps identified, plan in hand. All 14 engineers know where they’re strong and where to improve, so Rootstrap has a clear plan for what to train next.
  • Code review handed to agents. The assessment confirmed review as the biggest time sink, so engineers now hand repetitive review work to agents first.
  • Skills that work in any tool. The patterns apply in Codex, Cursor, and Claude Code alike.
  • A foundation to scale. Rootstrap can now standardize AI-native delivery across its broader engineering organization.

Win back your engineers’ time with AI agents

Eliza assesses how your team uses AI today and trains them on the agentic workflows that save the most time.

Contact us
→