Note: Pause. Let the silence work. Look around the room. Then click to the title.
an astrological essay for IT people and not only
Nikolai Ryzhikov · CTO, Health Samurai · Lisbon Agentic Meetup
Note: "To answer this, let me tell you a story about an industry that already faced exactly this question."
Note: Historical parallel - the core argument of the talk
Electronics was a distributed cottage industry
- ~150 TV manufacturers in the US alone (early 1950s)
- ~20,000 electronics repair shops
- ~1,500 component distributors
- 1.9M workers in electronics manufacturing
- 24,000 EE graduates/year at peak (1985)
Note: Paint the picture: every town had radio repair, local electronics manufacturers. 170 postwar TV manufacturers documented by name.
Hand-assembling discrete transistors & capacitors
Thousands of components → one chip
Per-unit cost: fractions of a cent
Note: The IC was the atom bomb of electronics manufacturing
| Disposable electronics | Barriers skyrocketed |
|---|---|
| Nobody repairs a $30 calculator | Fab cost: $4M (1970) → $20B (2020s) |
| Repair shops: 20K → 7K (-65%) | 5,000x increase |
| Entire maintenance ecosystem died | Small/mid manufacturers couldn't compete |
Note: Not one effect but two simultaneous destructions. Both killed the middle market.
Not gradual. Structural.
- US TV manufacturers: 150 → 0
- US DRAM producers: 14 (1970) → 3 (1986)
- Intel DRAM share: 83% (1974) → 1.3% (1984)
- Japan took 90% of 256K DRAM by 1984
- Electronics mfg jobs: 1.9M → 1.0M (-47%)
- US global fab share: 42% (1980) → 12% (2020)
Note: By 1988, Zenith was the sole surviving US TV manufacturer. Then it died too.
Niches too small or too demanding for mass production:
- Medical devices, aerospace, defense
- Fabless design companies (Qualcomm, Nvidia, AMD)
- Contract manufacturers for custom small runs
A tiny fraction of what existed before.
The survivors spend 21-28% of revenue on R&D
Note: Fabless share of global IC sales: 7% in 1999 to 35% in 2021.
Radio engineers manually designed & troubleshot circuits — that was routine intelligence.
Microelectronics automated it into chip design tools & fab processes.
The routine expertise became worthless.
Note: This is the thesis: the mechanism is making routine intelligence cheap. Same mechanism, different era.
Platforms replaced custom software development:
Salesforce · SAP · Oracle · Dynamics
A new middle market emerged: configuration & customization specialists
10-200 people firms. They don't build software — they adapt platforms.
Note: First disruption already happened. Custom dev shops became platform consultancies.
| $1.6T Global IT services market | 32% is config & customization |
| 47M Software developers worldwide | 15K+ Platform customization firms |
Note: Salesforce alone: 3,800 partners, 170K certified professionals, 9.3M related jobs. SAP: 3,460 partners, $16B consulting market.
Selling routine intelligence by the hour:
- India: 5.8M IT workers, $283B revenue, 55% of global outsourcing
- Ukraine: 300K+ IT specialists, 2,118 tech companies
- Global outsourcing market: $422-618B
- Upwork: 18M freelancers, 34% in software dev
cheap hands → cheap components → cheap AI
Note: Fiverr already cut 30% of staff to pivot to "AI-first". The signal is clear.
- Configure a CRM workflow? AI does it.
- Build a custom ERP report? AI does it.
- Set up permissions? AI does it.
Already happening:
- Salesforce Agentforce: -40% cost per interaction
- SAP Joule: -50% design iterations
- Gartner: 75% of new apps will be low-code by 2026
Note: 84% of enterprises already adopted low-code. The platforms are eating their own partners.
| -67% Entry-level dev jobs since 2022 | 245K Tech layoffs in 2025 |
| 57% US work hours automatable (McKinsey) | 44% hiring managers expect AI layoffs |
Fiverr cut 30% of staff to become "AI-first". Oracle cut 30,000.
Note: Companies that hired 5 juniors now hire 2 mid-level engineers with AI tools for same output.
| Electronics | AI |
|---|---|
| Building a fab: $100M+ | Training a frontier model: $1B+ |
Your 50-person Salesforce shop cannot build this infrastructure.
Note: Same dynamic: technology raises the barrier to entry beyond the reach of the current mid-market.
of semiconductor workforce is in R&D today
The routine work disappeared. Surviving companies need a far higher density of genuinely innovative engineers.
McKinsey: demand for higher cognitive skills to grow +19% through 2030.
But concentrated in large, well-capitalized companies.
Note: The middle market won't capture this growth. It concentrates at the top.
Democratization of Intelligence
| Before | After |
|---|---|
| Mainframes cost millions | Cheap chips made computing personal |
| Only corporations & governments | Entirely new industries emerged |
Note: Microelectronics didn't only destroy — it created the PC industry, far larger than what it replaced.
Expert intelligence has always been expensive:
- Legal counsel
- Financial analysis
- Medical diagnosis
AI makes routine expert intelligence cheap and universally accessible
Note: Just as mass production made cars available to everyone, AI makes professional intelligence available to everyone.
| TV (before) | YouTube (after) |
|---|---|
| Centralized | Millions of creators |
| Few networks controlled everything | on shared infrastructure |
AI does the same for software:
- Healthcare admin → AI billing assistant
- Lawyer → AI contract reviewer
- Construction PM → AI project scheduling
Note: Solo creators can't beat Salesforce on general CRM. But they can serve specific niches BETTER.
We cannot know the order
A: IT Middle Market Collapse
90% of configuration shops become non-viable. Outsourcing dies.
B: Platform Disruption
Salesforce, SAP, Oracle — built for expensive customization. AI-native apps deliver the same, cheaper.
C: The Open Source Wildcard
DeepSeek matched frontier performance at a fraction of cost. Capable models on your phone.
Note: We don't know which comes first. Maybe all three at once.
The current distribution of economic activity in software will not survive the decade intact.
Note: Pause here. Let it sink in.
Human + AI
Taylor's scientific management applied to AI:
- Specialized agents, rigid pipelines
- Orchestrators, DAGs
- Fully autonomous agent swarms
Note: This is what most people are building. And I think it's wrong.
"I haven't seen anything useful from fully autonomous agents yet — only human + agent."
The real breakthrough is not how autonomously an agent writes code — it's how well it works together with a human.
| Factory | Band (XP/Lean) |
|---|---|
| Specialized roles | Cross-functional team |
| Rigid pipelines | Flexible collaboration |
| Optimizing local tasks | Seeking global optimum |
| Known problems | Navigating uncertainty |
LLMs are generalists. Why force them into specialist roles?
Note: The architecture is ideologically loaded. Factory vs open space. I believe in the band.
XP's core insight: the feedback loop IS the product
idea → poc (agent) → try it → feedback
→ poc (agent) → try it → show
→ feedback → review
Instead of a spec — working code
The loop from idea to running code: → 0 minutes
But the loop still rotates around a human. Human as decision-maker. Agent as executor.
Note: Kent Beck wanted user stories tiny — as an excuse to talk. Now agents make the loop almost instant.
- Agree on architecture through dialogue, not a spec doc
- Tests become the spec — if they pass, skip reading the code
- Evaluating test quality < evaluating implementation quality
- Agents write both tests and code
- Humans control the tests
"A good test suite IS the best long-horizon task for an agent."
Note: Functional tests only. Cover user scenarios, not implementation. Rails got this right 20 years ago.
Instead of code review
- Multiple agents solve the same problem in parallel
- Each sees neighbors' solutions — can enrich their own
- Each defends their solution — why this, not that
- Final report: agreements & disagreements
Real result: Claude, Codex, Kimi on a session converter. Claude nailed it on medium. Codex failed on max after 3 extra prompts.
Note: DEMO opportunity here if time allows.
We used to build once — no energy to experiment.
Now: try 10 variants.
Write tests, say: "solve it with minimal code, zero deps"
"Simply making things is hard — making things simple is harder."
Note: Multi-sampling is the new TDD. Multiple solutions compared by machines.
Build fast with agents: less stack, not more
| Kill | Replace with |
|---|---|
| raw SQL | |
| htmx | |
| one file = one feature | |
| zero-dep single files |
"A business feature in one file with SQL, CSS and HTML — that's the stack. We reinvented PHP and it's beautiful."
Note: Every token in context has a cost: latency, money, attention.
We're building Agentic Workspaces
- A "room" where humans and agents collaborate
- Standard protocol (ACP) — plug in any agent, any client
- Skills, secrets, permissions per workspace
- Fork sessions for subtasks, merge results back
"The important qualitative leap: a chat room with several humans and possibly several agents."
Note: Not a factory. A jam session. Not configuring platforms — building vertical AI-native products.
- IT middle market is collapsing — same pattern as electronics
- Routine intelligence is being automated — 80-90% of current work
- New markets emerge — democratization, YouTube model
- Multiple futures, unknown order — all mechanisms already active
- Human + AI > autonomous AI — the loop rotates around humans
- Build, don't configure — domain expertise is your moat
- Flatten your stack — less complexity = better agents
Note: Summary slide. One takeaway per point, keep it fast.
The businesses that survived the electronics collapse were the ones that pivoted before the market collapsed, not after.
Note: End on urgency. The time to move is now.
Nikolai Ryzhikov · CTO, Health Samurai · @niquola
agentic-workspaces · consensus · hyper-code · agent-on-procs
Note: Open for questions. Mention the projects if people want to see real code.