In a column for O’Reilly Media’s Radar blog, Paul Iusztin argues that modern engineering requires shifting from trendy terminology toward structured software factories.
Iusztin, the founder of Decoding AI who spent more than seven years writing software before the coding-agent era, built his own software factory named Squid to ship small and medium projects with minimal human intervention, Stack Overflow reported. The author criticizes recent industry buzzwords like loop engineering and graph engineering for overexplaining intuitive concepts while ignoring actual development processes.
Instead, Iusztin frames automated development around the software factory model, drawing on discussions from the AI Engineer World’s Fair 2026. The workflow relies on an eight-stage architecture categorized into planning, building, and self-improving loops running over a shared context layer, Everything Blog reported.
The factory model divides the software development lifecycle into distinct functional buckets. The intake and brainstorming stages classify incoming work and research high-impact features using user data and technical research. Planning transforms research into documented tickets backed by Architecture Decision Record logs before deployment agents begin writing and reviewing code.
The actual building phase runs an implementation loop where software engineer and QA agents process tasks. Review and CI steps run test suites and check pull request diffs against product standards, triggering fixing agents automatically when failures occur. Staging releases require human deployment gates, ensuring engineers retain final authority before software ships.
Orthogonal to these steps sits the context layer, which prevents LLMs from guessing generic solutions. Iusztin utilizes file-based LLM wikis to supply structured agent memory from tools like Obsidian and Readwise.
The author recounts that his initial version of Squid chased full autonomy as a remote monolith, which became impossible to debug or halt mid-run when scripts failed. Engineers must retain control over high-impact strategic decisions during brainstorming and planning.
“The loop is the atom,”
Addy Osmani wrote, and a factory is
“an org chart made of loops.”
“Software engineering will become factory engineering,”
Zach Lloyd, CEO of Warp, said.
“My job is to write loops,”
Boris Cherny, who leads Claude Code at Anthropic, said.
“The whole loop, the whole lifecycle of developing software with autonomy,”
“Humans steer. Agents execute,”
OpenAI framed the dynamic after producing roughly 1,000,000 lines and 1,500 merged pull requests over five months with zero hand-written lines.
Iusztin founded Decoding AI and launched Squid three months before the article’s publication, initially as an overbuilt remote monolith that proved impossible to debug or halt mid-run. He then pivoted to local agents running in isolated codebases via worktrees, concluding that most features build sequentially and parallelization remains limited.
His workflow draws on file-based LLM wikis fed from Obsidian, Readwise, and Google Drive to supply structured agent memory, and he notes the bottleneck stubbornly stays at planning. He has also published a deep-dive on harness architecture titled “Building a Coding Agent From Scratch.”
To maintain stability, Iusztin argues that factories need granular commands for specific tasks alongside end-to-end capabilities for 24/7 autonomous graphs. Staging releases require human intervention to verify deployment checks before production integration.