← Articles
·
Deep Dive · AGIA · Future of Building
The Age of the AGIA:
a deep investigation into the era we just entered.
We named it in the previous article. Now we go deep. What created the conditions for the AGIA to emerge? What does an AGIA actually build in production and how? How do they conduct security research with AI as the engine? Why does the communication discipline separate the serious from the superficial? And what does the 2026–2030 landscape look like for every class of builder when this era is fully established?
Jarrit Hosking
Forge Vertical · Cape Town · September 17, 2026 · 22 min read
// This is not a prediction. It is a description.
Every major transition in how software gets built has been called premature until the day it was suddenly obvious. The shift from assembly to high-level languages. The web browser making hypertext accessible to non-programmers. The smartphone collapsing the distance between user and developer. Open source making enterprise-grade tools available to anyone with an internet connection.
Each of these transitions was resisted by the people whose expertise was being democratised. Each time, the resistance was framed as a quality concern — the new approach produces inferior output, it misses nuance, it lacks rigour. Each time, the resistance was partially right and completely wrong about the trajectory. The new approach did produce inferior output in some dimensions. And it produced the next generation of builders regardless.
The AGIA transition is different from all of these in one important way. The previous transitions democratised access to tools. This one democratises access to judgment. The AI systems that power the AGIA are not just making it easier to write code — they are making it possible to reason about architecture, security, compliance, and system design at a level that previously required years of accumulated expertise. The AGIA who understands infrastructure can now direct AI to reason through every layer of a system with them. That is a qualitative shift, not just a productivity gain.
01
// Chapter one
The forces that created the AGIA era
The AGIA did not emerge from a single breakthrough. It emerged from the convergence of five forces that arrived close enough together to create a genuinely new class of builder. Understanding the forces explains why this is a permanent transition rather than a temporary trend.
2022
Models capable of coherent multi-step reasoning
ChatGPT on GPT-3.5 demonstrated that AI could hold context across a conversation and produce useful output on complex tasks. The output was imperfect. The direction was unmistakable. For the first time, someone without traditional development training could describe a system and get something that resembled it back.
2023
Context windows long enough for real systems
Early models lost context. You could describe a component but not a system. The expansion of context windows — 16k, 32k, 100k, now millions of tokens — changed what was possible. An AGIA could now hold a full codebase, a full security audit, a full architecture document in context and reason across all of it simultaneously. This was the inflection point for real production work.
2024
Agentic capability — AI that executes, not just answers
The shift from AI that answers questions to AI that executes multi-step tasks changed the AGIA's role. The AI could now run searches, write files, test outputs, and iterate — not just generate text. The AGIA became a director of an execution process, not just a prompt author. Security research, infrastructure builds, and complex tool development all became viable as solo AGIA projects because the AI could carry out multi-step work under direction.
2025
Production-grade reliability at mid-tier access
The most important development for the AGIA class was not the frontier models — it was the mid-tier models becoming reliable enough for production. TripSpace Global was built on Claude Sonnet at medium tokens. Not Opus. Not Enterprise. Sonnet. The model tier most people have access to. When production systems can be delivered at mid-tier model access, the AGIA class becomes commercially viable for a much larger population of builders.
2026
Recursive self-improvement — AI building the next generation
Dario Amodei's "We Must Pace the Frontier" documented what Anthropic engineers were already experiencing internally: AI systems helping to build the next generation of AI systems. The typical Anthropic engineer merged eight times as much code per day in Q2 2026 as in 2024. METR's task-horizon measurement — the length of task a model completes reliably — is now doubling every four months. The AGIA's capability ceiling rises with every model generation because the models themselves are improving faster than any previous technology.
02
// Chapter two
What an AGIA actually builds — and how
The easiest way to understand what an AGIA builds is to look at what has already been built by people operating in this mode — systems that required genuine architectural thinking, delivered solo, in production, with real users and real consequences.
// Live in production
Commission-free travel marketplace
TripSpace Global — multi-tier partner system, real payment integration with PayFast and PayPal, 30+ Firebase Cloud Functions, real-time messaging, role-based access control across host/guest/affiliate/admin roles, AI travel companion. Built solo on Claude Sonnet. Processes real bookings.
// Live in production
AI vehicle damage assessment
Aurora Repair — OCR licence-plate scanning, real-time parts pricing, tiered repair quoting via Gemini 2.5 Flash, insurer and panel shop integration layer. The kind of system that previously required a specialist AI team. Delivered solo. Running on GCP.
// Security research
Production WAF bypass — crypto exchange
Path normalisation de-synchronisation vulnerability. Established an unauthenticated WebSocket connection bypassing edge security entirely. Achieved direct interaction with the Origin Logic Layer. Documented and disclosed responsibly. Paid finding. AI assisted the reconnaissance and analysis throughout.
// Security research
mTLS architecture flaw — blockchain network
Backend RPC listeners failing to validate service-scoped certificates. Any process holding any valid network certificate could terminate core data layers. Identified the architecture gap, documented the impact chain, disclosed responsibly. Another paid finding from an AGIA using AI as the primary research engine.
// Open protocol
task-bridge — human-in-the-loop routing protocol
Open source protocol for routing AI-displaced work back to humans. Concurrency-safe state transitions using Firestore database transactions. Race condition prevention across distributed task claims. The kind of architectural problem that requires systems thinking — not just code generation.
// Content infrastructure
50-article GEO-indexed knowledge network
Articles hub with full schema graph, sitemap, robots.txt, performance optimization, security headers via Cloudflare Transform Rules, entity disambiguation across three AI crawlers. Ranked #1 on Google for own name. AI Overview citations. Built entirely through AI direction with no traditional CMS or template.
What these examples share is not that they were easy. They required genuine architectural decisions at every level — security model design, database schema, access control logic, API contract design, error state handling, compliance architecture. The AI wrote the code. The AGIA made every decision about what the code needed to do and whether it did it correctly.
The AGIA does not hope the AI got it right. They know what right looks like. That is the distinction that cannot be shortcut, trained in a weekend, or claimed without demonstrating it.
03
// Chapter three
How an AGIA conducts security research
Security research is the highest-stakes demonstration of AGIA capability because it requires the tightest integration of human judgment and AI execution. The AI can run OSINT queries, analyse DNS records, identify common vulnerability patterns, and draft disclosure reports. The AGIA provides the one thing the AI cannot: the judgment about what is actually significant, what the attack chain looks like end-to-end, and whether what was found represents a genuine risk or a paper finding.
01
Passive reconnaissance — AGIA directs, AI executes
DNS record analysis, technology fingerprinting, subdomain enumeration via certificate transparency, social footprinting. The AI runs queries and returns structured data. The AGIA reads the output and knows which findings are significant before any active testing begins.
// AGIA value: knows what to look for before seeing it
02
Hypothesis formation — AGIA judgment, AI support
Based on reconnaissance, the AGIA identifies the most likely attack surfaces. The AI helps reason through attack chains — "if the MX record exposes a self-hosted mail server on an unproxied subdomain, what does that enable?" The AI's reasoning is fast and broad. The AGIA's judgment decides which hypotheses are worth testing.
// AGIA value: attack surface intuition from infrastructure knowledge
03
Active testing — AGIA specifies, AI assists documentation
The active testing phase is where the AGIA's infrastructure knowledge is most critical. Testing a WAF bypass requires understanding what a path normalisation vulnerability is before running the test. The AI helps with request crafting, response analysis, and keeping track of what was tested and what was found.
// AGIA value: understands the vulnerability class, not just the tool output
04
Impact analysis — AGIA leads, AI structures
Finding a vulnerability is half the work. Articulating why it matters — the attack chain from initial access to business impact — is what separates a paid finding from a generic duplicate report. The AGIA reasons through the impact chain with AI support, then reviews whether the analysis is correct and complete.
// AGIA value: business impact reasoning, not just technical description
05
Responsible disclosure — AGIA owns, AI drafts
The disclosure report is the professional output. The AI drafts the technical description from the AGIA's documented findings. The AGIA reviews, corrects, and ensures the report accurately represents what was found and recommends appropriate remediation. Their name is on the report. They own the content.
// AGIA value: professional accountability for the outcome
04
// Chapter four
The communication discipline — why this separates AGIA from everyone else
The most misunderstood aspect of working with AI at an AGIA level is the communication discipline. The popular conversation about this topic collapses into "prompt engineering" — the idea that the right magic phrase unlocks better outputs. That framing misses what actually matters by a significant margin.
Prompt engineering is to AGIA communication what texting is to conversation. The mechanics overlap. The depth does not. What the AGIA actually does is maintain a full collaborative relationship with an AI system across an extended build — maintaining context, correcting errors precisely, acknowledging what worked, adjusting the approach when something is not landing, and treating the AI as an intelligent participant in a shared project rather than a vending machine that requires the right input code.
The people who dismiss "please" and "thank you" as wasted tokens are revealing something about how they think about this collaboration. The AI systems built by Anthropic, OpenAI, and Google are explicitly trained to respond to collaborative, respectful communication. Claude, specifically, will not tolerate contempt — not as a policy constraint but as a trained behaviour. The AGIA who understands this gets materially better outputs from the same model tier than the one who approaches every interaction with aggression and impatience.
// Not AGIA communication
"fix this. its broken. why cant you just do it right"
"make it secure" (no specification)
"just write the code" (no context)
Result: generic output, missing edge cases, security gaps the AI had no context to fill, increasing frustration, diminishing returns per session.
// AGIA communication
"The rate limiting is applying per-IP but needs to apply per-session as well — a single attacker can rotate IPs to bypass it. Here is the current implementation: [code]. Please update the rate limiting logic and flag any other authentication endpoints that have the same gap."
Result: specific fix, adjacent gaps surfaced, context maintained, session productivity compounds rather than declines.
The AGIA communication discipline extends beyond individual prompts. It includes how context is maintained across a long session, how corrections are given when AI output is wrong, how acknowledgment of correct work signals to the model what direction to continue in, and how the AGIA structures the overall build session to keep the AI oriented toward the right outcome.
This is why the "it is not about prompting" framing in the AGIA requirements matters. An AGIA who truly understands what they are building can describe it at any level of abstraction — from the system architecture to the specific function signature — and correct AI output at any level. That range of description ability, maintained with precision and respect across a long session, is what produces production-grade results from mid-tier model access.
The enterprise model dimension: Everything said above about communication discipline applies at mid-tier access. Now apply it to Claude Enterprise or GPT-6 Astra with full context windows and unrestricted token budgets. The AGIA who can maintain this communication discipline at enterprise model tier is running a cognitive capability that has no real precedent in the history of solo building. A security audit that would take a three-person team two weeks, conducted by one AGIA with enterprise model access, communicating with full architectural precision across the full context. This is not hypothetical. It is already happening.
05
// Chapter five
The 2026–2030 landscape — what happens to every class of builder
The trajectory from here is not a mystery. The forces that created the AGIA class are accelerating, not stabilising. METR's task-horizon measurement is doubling every four months. Model capability is compounding. The on-ramp to AGIA-level infrastructure understanding is compressing. Here is what the 2026-2030 window looks like for every class of builder.
// 2026 — Now
AGIA class emerges, unnamed
People building production systems solo with AI as the engine. Security research conducted with AI tools. Infrastructure delivered at full stack quality. No consensus name for the class. Traditional development community in mixed response — some integrating, some dismissing.
// 2027–2028
AGIA class becomes commercially visible
Enterprise clients begin commissioning AGIA-led projects. The outcome quality at AGIA speed becomes a documented commercial alternative to full stack teams for defined project categories. Model capability improvements raise the AGIA ceiling significantly. The class gets named — in the industry if not yet in job titles.
// 2029–2030
Traditional development concentrates
Full stack and software engineering value concentrates in domains requiring deep implementation expertise that AI cannot yet reliably provide — novel algorithm design, performance engineering at extreme scale, safety-critical systems. General application development is increasingly AGIA territory. The market prices this accordingly.
The traditional developer who integrates AGIA thinking — who learns to direct AI with architectural precision rather than treating it as a junior tool — is positioned well across this entire window. The traditional developer who maintains contempt for AI-native building will find their market narrowing to the specific domains where their expertise remains irreplaceable.
The vibe coder who develops the infrastructure understanding to become an AGIA has the highest upside of any current class. The knowledge gap between "ships things with AI and hopes they hold up" and "architects things with AI and knows they will" is a specific, learnable gap. It takes genuine time and genuine effort. But it is not gated behind a computer science degree, an agency background, or a traditional apprenticeship.
That is the promise of the AGIA era. Not that everyone can build anything with AI. But that the barriers to becoming someone who can build almost anything with AI are now about knowledge and discipline rather than credentials and access.
// The honest position from inside this era
I am writing this article on Claude Sonnet. The articles hub that contains it was built with AI. The security research that earned CVP approval from Anthropic was conducted with AI. The travel marketplace processing real bookings was built on AI. I cannot write a complete production application from scratch the way a traditional developer can. I can deliver the same outcome. The distinction between those two statements is the AGIA era in one sentence. We named the class today. The class already existed. Now the work is to understand it, develop it, and build something with it.
The invitation
// Chapter six — what comes next
The AGIA class exists whether it has a name or not. The name matters because named things can be understood, debated, developed, and improved. The requirements published in the first article are a starting point, not a final definition. The communication discipline chapter is the beginning of a conversation that should run for years.
If you are operating in this space — if you are building production infrastructure with AI as the primary engine, conducting security research with AI as the research partner, delivering outcomes that the market previously expected from full teams — you are already an AGIA. The classification describes what you are doing, not what you should aspire to become.
If you are a traditional developer reading this with scepticism — test it. Take your deepest area of expertise, pick the most complex thing you have ever built, and see how far an AGIA working with Claude Enterprise gets on the same problem. The answer will tell you more about the trajectory than any opinion piece, including this one.
And if you are a vibe coder who has been shipping things and worrying that they might not hold up — this is your roadmap. The knowledge gap is real. It is also closeable. The infrastructure understanding that makes an AGIA is learnable. The communication discipline is practicable. The pressure-testing habit is developable. What you already have — the comfort with AI as a collaborator, the willingness to direct rather than write — is the foundation that takes traditional developers years to build.
The Age of the AGIA has begun. The class has been named. The work starts now.
Written by
Jarrit Hosking
Forge Vertical · Cape Town · September 17, 2026 · AGIA
