Have agents write, run and fix their own code
PwC rebuilt its software work with CrewAI agents that write code, run it and fix their own errors; a PwC leader says accuracy rose from roughly 10% to 70%+.
Done with CrewAI
What they did
PwC needed faster, more accurate code in a programming language of its own, plus long planning documents. Its early AI tests got no feedback on whether the code worked. The team rebuilt the process on CrewAI. Agents write code, run it, read the error records and fix it in repeated passes. Agents also draft and refine planning documents while consultants give feedback. CrewAI's tracking showed leaders the agents' choices, run times and accuracy.
How it went
PwC's Jacob Wilson is quoted saying accuracy went from roughly 10% to 70%+ once CrewAI agents joined the work. The case study also says complex documents got done faster, without giving a figure.
Worth knowing
This is CrewAI's own case study. PwC says seeing each agent's task times, tool choices and share of work versus humans was key to proving the payoff and winning back consultants' trust.
Try it yourself with CrewAI
Set up a CrewAI crew that writes code in [language] for [task description]. One agent writes the code. A second agent runs it in [test environment] and reads the logs and errors. If it fails, send the log analysis back to the first agent for a fix, for up to [N] attempts. Record each attempt's outcome and runtime in a summary table. Stop and show me the final code and test results. Do not deploy or merge anything without my approval.
Swap the highlighted [parts] for your own details.
Source: CrewAI case study
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
