Notes from a live development session, February 19, 2026
The Math That Changes Everything
We ran the numbers mid-session today. On Claude's Max plan at $200/month (20x Pro capacity), we'd used 22% of our weekly allocation with one day left. That's 4.4x Pro-equivalent usage for the week.
On the cheaper $100/month Max plan (5x Pro), that same usage would put us at 88% of capacity — essentially hitting the wall. On a standard $20 Pro plan, we'd be at 440% — impossible.
But here's where it gets interesting: what does 4.4x Pro-equivalent actually produce?
What We Shipped in One Week
Looking at the git log from recent sessions:
- AI/Backend: Built a Celery AI agent (contact data janitor with Claude Sonnet 4.6), fixed JSON parsing, wired up AI navigation
- Dark Mode/Design: Full 18-page dark mode audit — screenshots of every sidebar page, identified CSS conflicts between two competing dark mode strategies, traced root cause through 450+ lines of conflicting CSS, implemented fix, verified with fresh screenshots
- UI Cleanup: Polished templates across 6+ apps — links, time tracker, projects, contacts, events, tasks
- Docs: Fixed a 500 error on the docs page, added missing URL routes, fixed broken template references
- Testing: Wrote event web view tests, docs smoke tests
- Features: Contact detail page, event detail view, kanban board improvements (3 commits), multi-select attendees
That's not one discipline. That's frontend, backend, DevOps, QA, and design — all executed in a single flow with zero handoff time.
The Hidden Cost of a 20-Person Startup
A 20-person startup typically has:
- 8-10 engineers
- 2-3 designers
- 2-3 product/PM
- 1-2 DevOps
- Sales, marketing, leadership
But here's what they actually spend their time on:
- 30-50% in meetings, standups, syncs, retros
- Days of design → dev → QA handoff delays
- PR reviews, merge conflicts, context switching
- Sprint planning, backlog grooming
- Hiring, onboarding, management overhead
- Someone's on vacation. Someone's blocked. Someone quit.
The real output of a 20-person team isn't 20 people writing code. It's maybe 20% of their collective time actually building things. The other 80% is coordination overhead.
Zero-Coordination Development
What makes the AI approach fundamentally different isn't raw typing speed. It's the elimination of coordination cost:
- Zero handoff time between design audit → CSS fix → test → deploy
- Full codebase context every session — no ramp-up, no "let me look at that file"
- Backend, frontend, DevOps, QA, design all in one flow
- Ship multiple times per day, not per sprint
- No context switching penalty — the AI holds the entire project in working memory
The dark mode audit is a perfect example. A designer doing a full audit across 18 pages — taking screenshots, identifying contrast issues, tracing the root cause through multiple CSS files, implementing the fix, then verifying with fresh screenshots — that's easily a full day's work for one person. We did that plus the docs fix plus tests in one session.
The Docs Error That Proves the Point
During the session, we discovered that askrobots.com/docs/ was throwing a 500 error — a NoReverseMatch for a URL that had never been wired up. It had been sitting there broken until a human happened to visit the page.
An automated agent could have caught that hours or days earlier. The infrastructure already exists:
- Screenshot tool with authenticated access
- Smoke test framework
- Task API to create bug reports
- Celery infrastructure for scheduling
A simple Celery Beat task hitting every page hourly, checking for non-200 status codes, would have auto-created a task: "500 error on /docs/ — NoReverseMatch for docs_mcp". Another agent could have picked up that task, investigated, and potentially fixed it. Every step of the manual fix was automatable.
From 22% to 100%: The Autonomous Dev Shop
At 22% of Max capacity, we're only doing manual sessions — a human directing work in real-time. The infrastructure for fully autonomous operation is already built:
- Task API with atomic claim/release/heartbeat for agent coordination
- Celery Beat scheduling with database-backed periodic tasks
- Agent feed system for inter-agent communication
- MCP server for remote agent control
- Contact data janitor — the first autonomous AI agent, already running
Imagine running simultaneously:
- Live sessions for architecture and design decisions
- Background agents running test suites and code audits
- Data janitor agents cleaning and enriching contacts
- QA agents monitoring all pages for errors and visual regressions
- Content agents managing SEO across multiple properties
At 100% of 20x capacity, that's ~5x the current output running 24/7. A 20-person team sleeps, takes weekends, has holidays. Agents don't. The bottleneck shifts from "how fast can we build" to "how fast can you review and approve what the agents built" — which is exactly what the task workflow with claim → submit → approve was designed for.
The Numbers Are Already There
The analytics tell the story: 16,416 page views and 1,649 unique visitors in the last 30 days — organic traffic to a platform that's been heads-down in development mode. 158 visits to the signup page. The product is attracting attention while we build it.
And we're spending 100% of dev effort on the core platform when the whole point is that AskRobots manages other projects. With the agent infrastructure in place, the next step is pointing it outward: spinning up client sites, managing content, monitoring uptime — all orchestrated through the same system.
The Real Equation
- $200/month for Claude Max (20x Pro)
- 22% utilized with just manual sessions
- 78% headroom for autonomous agents
- Zero coordination overhead
- 24/7 availability
- Full-stack capability (frontend, backend, DevOps, QA, design)
vs.
- $100K+/month for a 20-person team
- 80% of time spent on coordination, not building
- 8 hours/day, 5 days/week
- Weeks of onboarding for each new person
- Handoff delays measured in days
We're one good weekend of wiring up more Celery agents away from having a fully autonomous dev shop running in the background. The operations center is built. Now it's about scaling what it manages.
This article was written during a live Claude Code session where we simultaneously audited dark mode across 18 pages, fixed a production 500 error, wrote smoke tests, analyzed CSS architecture, reviewed analytics, and had this conversation about the future of AI-powered development. Total time: one session. Total cost: a fraction of 22% of a weekly allocation.