Skip to content
2026ContactGet updates
All ten talk summaries

SDLCAI 2026 / Talk summary

Muhammad Waseem

Code Was Never the Hard Part. Understanding Was—and Still Is.

Muhammad Waseem argues that AI makes code generation easier while understanding, integration and ownership remain difficult. He examines repository context, decision rationale and agent-facing applications, then proposes deterministic checks, human judgement and gates for safety, architectural fit and measurable value.

30:54 · 13 October 2026

Watch full recordingProgramme & current materials
  • system understanding
  • code review
  • repository context
  • decision records
  • agent memory
  • replaceable architecture
  • verification

Takeaway graphic

Share three selected takeaways from this talk. Download a landscape or portrait card, with suggested post copy and image alt text.

Understanding remains the hard part. Muhammad Waseem at SDLCAI 2026. Three paraphrased talk takeaways: 1. Check whether a change preserves the wider system. 2. Record decision reasons, constraints and rejected options. 3. Combine deterministic checks with human judgement and value gates. sdlcai.org
Three paraphrased takeaways, with timestamps from the final recording.
Landscape PNG · 1200 x 630Portrait PNG · 1080 x 1350Download post copy & alt text
Read post copy & alt text

Suggested post copy

Understanding remains the hard part Three takeaways from Muhammad Waseem at SDLCAI 2026: 1. Check whether a change preserves the wider system. 2. Record decision reasons, constraints and rejected options. 3. Combine deterministic checks with human judgement and value gates. Paraphrased from “Code Was Never the Hard Part. Understanding Was—and Still Is.”. Watch the full talk: https://www.youtube.com/watch?v=o_WaA-EaBqo #SDLCAI

Image alt text

Understanding remains the hard part. Muhammad Waseem at SDLCAI 2026. Three paraphrased talk takeaways: 1. Check whether a change preserves the wider system. 2. Record decision reasons, constraints and rejected options. 3. Combine deterministic checks with human judgement and value gates. sdlcai.org

Key points

Paraphrased from the talk. Follow each timestamp to hear the argument in context.

  1. Measure whether generated changes fit and preserve the system, rather than judging a file or snippet in isolation.

    7:09–11:01

  2. Review and understanding have to keep pace with generated work; more code alone does not remove that constraint.

    11:01–14:58

  3. Decision rationale, rejected options, constraints and agreement need to be recorded because code does not contain all organisational knowledge.

    14:59–16:58

  4. Design for replacement of models, tools and agents while retaining ownership of the data and associated knowledge.

    21:11–22:26

  5. Use deterministic checks, human judgement and explicit gates for safety, architectural grounding and actual value.

    22:26–24:07

The talk in context

Muhammad Waseem organises the talk around writing, reading, deciding and caring for software. Faster generation expands the scope in which mistakes can propagate, while file-level success can conceal broken relationships elsewhere in a repository. His account of a Java refactoring study motivates attention to dependencies and system-level evaluation. He then examines how review capacity, limited agent memory and unwritten organisational knowledge constrain adoption. Decision reasons, rejected options and ownership are not fully represented by source code. Looking ahead to applications whose users include other agents, he discusses intent, contracts and the ability to replace models, tools and agents while retaining control of data and knowledge. His proposed response combines deterministic checks with human evaluation of consequential cases, and asks whether a change is safe, grounded in its environment and producing the intended value.

Continue with another talk

Explore related questions across the seminar.