Contents
In brief
In a large monolith the bottleneck is rarely that developers type slowly. A hypothesis crawls through approvals, backlog, design, review, and release—and the opportunity window closes first. Handing everyone an AI coding assistant speeds one segment of a long queue, not the whole production cycle.
Ilya Radchenko (SimpleOne) on Habr puts it bluntly: agentic development changes how work is executed—roles split across agents—but without infrastructure it is an expensive toy. You need a tracker, an execution environment, data access, and a tool layer such as MCP; a low-code platform often becomes a practical substrate.
What happened
The classic path keeps a human as the single choke point: clarify requirements, write code, check side effects, run review. In an agentic path roles appear—requirements, specification, plan, code, review, smoke checks, build—while the human sets boundaries and confirms the outcome still solves the original problem.
The author lists a minimum foundation. Without a shared board of statuses, an “orchestra” of agents becomes chat chaos. Without an IDE, runtime, and version control you are stuck copying from a dialogue into a repo. Without application data an agent cannot prioritize a backlog or see architectural impact. Without standardized tool access (MCP over APIs) every action is costly and fragile.
Low-code here is not a “no code” slogan but an environment that already holds the data model, processes, UI, and runtime. The bridge between simple scenarios and pro-code cuts friction: agents navigate one contour that remembers platform entities.
Why it matters
Companies often score AI success by minutes to the first diff. In a monolith the value is time from hypothesis to a production check. If the rest of the pipeline is unchanged, faster typing only lengthens the review and regression queue.
SimpleOne’s cases show order-of-magnitude gaps. An upstream/downstream flow dashboard: classic estimates in weeks and person-months; with agents—days of iteration and dozens of human checkpoints. Multi-factor backlog scoring across CRM, ITSM, and SDLC with auto-recalc over thousands of records tells the same story: months for a team versus a few working days on the happy path.
A paired “skeptic” agent that rejects drafts (over half in their project logs) catches contradictions with business rules, not only syntax. Quality gates stay in the loop even when generation is faster.
In practice
- Map friction first (approvals, context switching, release)—not a shopping list of models.
- Give agents a shared task board with statuses and explicit human confirmation points.
- Wire IDE, runtime, and VCS so artifacts land in the repository, not the clipboard.
- Expose application data as described tools (MCP), not a pile of Confluence links.
- Price friction and interface errors, not tokens alone: cheaper often means fewer handoffs.
An experienced human in the loop is mandatory: a novice with one prompt will not get a mature solution in a hard domain. Rules, algorithms, and required checks are design, not optional polish.
Takeaway
Assistants speed typing; agents can reshape the SDLC—but only on a platform foundation: tasks, environment, data, integrations. A low-code surface with rich context lowers the cost of assembling that contour. The question is less “will you trust agents with everything?” than where you deploy them so value is not eaten by friction and token bills.