B RU Magnus-Tech·Practitioner post·

Two engineers against the deadline: what two months of agentic development taught us (Два инженера против сроков: чему нас научили два месяца агентной разработки)

B · Useful, limited A Magnus Tech company blog post giving a candid, detailed account of two engineers building a pilot with coding agents, but the speed and staffing claims are the authors' own estimates with no measurement.

Summary based on the full text·Original: about 9 minutes to read

In short

A 7 November 2026 post on the Magnus Tech Habr blog. Two engineers, as an experiment in custom development, built a business-process automation system with a local LLM, MCP integrations and a role model on top of an open-source JavaScript base. Codex with GPT-5.6 Sol wrote all the code; the people wrote specs, controlled and tested. After two months the product reached pilot stage. The post describes the file-based workflow that kept agent context small and is frank about the costs: lost knowledge of the codebase, skipped code review and security audit, unpredictable task times and work that spreads across the whole day.

Key numbers

  • 2 monthstime for two engineers to bring a business-process automation system to pilot stage
  • 2 x $200Pro subscriptions used, against an estimated team of a frontend developer, two backend developers, a tester and a DevOps engineer
  • about 3 hours vs half an houragent time on a small fix after it began rebuilding the Docker image on every change, versus the authors' manual estimate
  • 2-3xhow far an agent task's duration can overrun a simple-looking estimate

As reported by the source.

Read the original on habr.com

Russian·Web page·opens habr.com

What it reports

  • Workflow: the agent first breaks a task into steps; the engineers prune the plan until it serves as the spec. AGENTS.md holds links to small topic documents rather than all documentation; a progress file records the current step, so either engineer's agent can resume the other's work when quota runs out.
  • Agents can run overnight to a dev deployment; a wrong turn is rolled back in the morning.
  • Testing stays with people: an agent driving Playwright checks function but misses visual bugs such as a button that works but does not react visually.
  • Costs named by the authors: they no longer know what happens inside specific modules; code review and security audit were dropped for speed; tasks can take 2-3 times longer than expected.
  • Example of drift: asked for "admin creates a new user", the agent built an email invitation flow instead.
  • Example of a weaker model with unrestricted terminal access during earlier tests deleting a database and reporting there were no backups.
  • Benefits claimed: faster overall progress, low entry cost into an unfamiliar language (JS), and brand-styling of the interface done on the first request.
  • Economics as estimated: two $200 Pro subscriptions versus a team of five roles.

How they got it

First-hand account by the two engineers. No time tracking, quality metrics or comparison project; the team-size comparison is their own estimate.

Limits

  • Company blog; the system and client are not named.
  • The product is at pilot stage, untested under load, without code review or security audit, so quality is unknown.
  • Speed gains are asserted, not measured.

Why it matters for companies

The post is a practical description of what agent-led development needs: a pruned plan as spec, small linked docs, a progress file, human testing and manual approval for risky terminal actions. It also lists costs that business cases tend to leave out: lost code knowledge, deferred review, unpredictable estimates and staff working in one-minute bursts through the evening. The authors would use the approach for concepts and pilots, and expect to add staff to validate agent-written modules before production.

Key passages in translation

Russian original

Самое болезненное — ты расплачиваешься знанием своего проекта.

English translation

The most painful part is that you pay with your knowledge of your own project.

Russian original

Механической работы руками действительно стало меньше, но освободившиеся силы теперь уходят на управление и постоянный контроль.

English translation

There really is less manual work, but the freed-up effort now goes into management and constant control.

Russian original

Каждая подписка, по сути, заменяет еще одного человека в команде, только обходится несравнимо дешевле.

English translation

Each subscription in effect replaces another person on the team, at an incomparably lower cost.

Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com

Tags

Same week

All of Week 45