AI-assisted software development

created: Mon, 22 Dec 2025 03:03:54 GMT

Recently we've developed a software project using our agentic problem solving framework, one of a kind.

  • agentic problem solving framework:
    • problem statement and validation agents
    • solution proposer and reviewer agents
    • solution implementer and tester agents
    • documentation and follow up agent
  • Project stats: 30k sloc, 500 issues, 3000 tests, very well documented
  • All decisions are documented as problems with solutions and a part of the project itself
  • Established agentic development flow allows continuity:
    • create new problem to solve
    • solve the problem with the framework

Reflections

  • it's easy to develop a project that hits all the marks (unit test coverage, documentation, etc)

  • it's still may fail functionally, but it's easy to capture and fix

  • functional and e2e-tests are extremely important, but slow

    • requires engineering to speed it up

  • user acceptance tests are important

    • requires process engineering to capture feedback and inject it into theh flow

  • no jira or confluence is required:

    • develop a dashboard takes 30 minutes
    • rag/grep + ai to navigate

  • users with no software development background are capable of making changes

    • different kind of changes:
      • easy to implement within the current architecture (agents do very well)
      • requires architectural changes (agents require humand guidance and intervention)
    • these changes are easy to spot after the implementation

  • possible to implement the same problem multiple times to pick up/converge on solutions

  • very valid question still stands: how do we know that software works, with the same answer

    • documentaion and unit tests
    • automtated e2e and acceptance tests

Detour

  • good old software development practices aka agile and extreme programming are applicable directly
    • agile manifesto
      • individuals and interactions over processes and tools
      • working software over comprehensive documentation
      • customer collaboration over contract negotiation
      • responding to change over following a plan
    • extreme programming
      • user stories are written
      • make frequent small releases
      • the project is divided into iterations
      • iteration planning starts each iteration
      • refactor whenever and wherever possible
      • create spike solutions to reduce risk
      • no functionality is added early
      • code must be written to agreed standards (skills!)
      • code the unit test first
      • all code must have unit tests
      • all code must pass all unit tests before it can be released.
      • when a bug is found tests are created
      • acceptance tests are run often and the score is published

Synthesis

  • generic agentic problem solving framework

    • define intent and constraints and let the system figure the rest out iteratively in small steps
    • engineer requirements and constraints validation
    • develop transferrable skills

  • same framework is applicable to different stages of software development

  • current software processes and roles are tailored to reduce risk and dependency on individuals

    • scrum, jira and so forth

  • the change is scary and rightfully so, because ai is a process amplifier

    • amplified shitty processes produces way more shit
      • examples of developers pushing ai slop
    • should empowering individuals instead (agile!)
    • customers could and should make changes directly

Predictions

Agents and LLM increase productivity, by now this is an established fact. Another established fact is that this productivity is not always a good thing (ai slop).

The next year would be a massive process change to accommodate increased productivity.