Ansai Docs mark Ansai Docs Cornerstone documents of Ansai Technologies

← All documents

Dira — Entry 02: The builders' playbook (AI-native & agentic-native infrastructure)

Recorded 2026-09-24. This entry is the "how" companion to Entry 01's "what/why". Distilled from deep research across engineering writers, AI data startups, developer platforms, and Africa-first building realities. Sources are named inline; open questions are marked plainly.

1. What the best engineering writers agree on

Gergely Orosz (The Pragmatic Engineer) — "Slow down to speed up." Agents changed engineering faster than leadership adapted: 5x PRs, 2.5x code, 3x larger PRs vs 18 months ago; Anthropic runs 70–90% Claude-generated code. Individual productivity is up, team productivity is flat, quality is dropping. His core claim: verification is the scarce resource, not generation. Review the plan/tests/schema, not the implementation; triage reviews by blast radius (the OpenAI/Anthropic approach). Prediction worth noting: CI/CD flips from speed to risk-reduction — more, slower tests.

Petar Ivanov (The T-Shaped Dev) — "An agent is mostly not a model." The model is a probabilistic core inside a deterministic machine, and you engineer the machine: the 3-layer testing pyramid — deterministic unit tests for the substrate, LLM-judge scenario tests, continuous evals in production. Never mock the model and assert on the mock.

ByteByteGo — pattern-first: RAG vs agents is a decision rule (RAG = one retrieval + one generation: cheap, predictable, debuggable; agents = reasoning loop with tools: use only when the task needs it). For production: observability from day one, split read vs write operations into different architectures, and put an MCP gateway in front of tools (credential vaulting, allowlists, human-in-the-loop approvals, audit trail).

Where they converge: production AI is systems engineering, not prompting. Humans stay in the loop on high-risk actions. Separate the concerns: the model, the machine around it, and the durability layer. Pick the cheapest pattern that works.

2. The money trail: AI data startups that exploded

The market is screaming one thing: data for agents is where the money is.

The data flywheel, stated plainly: capture → label → train → deploy → capture. Whoever owns the capture point in a vertical owns the flywheel. Our capture points are WhatsApp conversations, school operations data, SACCO transactions — African verticals the Valley isn't watching.

Africa's own trail: InstaDeep → BioNTech (~$550M, completed Jan 2023); Aerobotics $26M+; DataProphet $16M; Intella (Egypt) $16.9M; Sama (Kenya) ~$85M. African AI startups raised $1.25B between Jan 2019 and Mar 2025. But: AI-native startups got <2% of African funding in H1 2026 — the money flows to applied AI. Lesson: sell the outcome (the school runs better), not the technology (the agent stack).

3. The three developer platforms: how we use each

Anthropic (Claude) — the agent builder's platform. Claude API with Opus 5.5 ($4 input / $20 output per million tokens), Sonnet 5 ($2/$10), Haiku 4.5 ($1/$5). The real assets for us: Agent Skills (open standard since Dec 2025 — SKILL.md + scripts + references, progressive disclosure; works across Claude Code, Cursor, OpenClaw), MCP (donated to the Agentic AI Foundation; 97M downloads, 10,000+ enterprise servers), and 1M-token context on the flagship. Use Claude where agent reliability matters most — the orchestrator, the reviewer, the mother agent.

Google (Gemini) — the cost-efficient workhorse. Gemini 3.6 Flash at $0.75/$3.75 per million tokens (intro pricing through Dec 2026, then $1.50/$7.50); 2.5 Flash around $0.30/$2.50. AI Studio has a free tier; Vertex AI is the enterprise path (no perpetual free tier, but zero-training-data-retention by default — matters for client data). Agent Development Kit (ADK) is open-source with native A2A support; Gemini CLI is free (1,000 requests/day). Use Gemini for high-volume worker tasks — classification, extraction, drafting.

Meta — the channel, not (yet) the model. WhatsApp Business Cloud API: access itself is free, no paid middleman required. Critical, dated fact: from Oct 1, 2026, service messages inside the 24-hour window — previously free — become charged at ~$0.004 (about KSh 0.52) per message in Kenya. Every WhatsApp agent we ship must be designed around this pricing from day one: batch, be concise, resolve in-window. (Meta's model API details we could not verify this run — treat Llama as the self-host weights story: Ollama/vLLM for local inference.)

The multi-model gateway: never call providers directly from product code. OpenRouter (one key, 400+ models, free tier) or self-hosted LiteLLM (MIT, your keys, fallbacks/cooldowns/retries) in front of everything. Route by task: cheap model for classification, strong model for reasoning, fallback across providers when one is down. This is also our hedge against any single lab changing prices or terms.

4. Building in Africa: the nuanced playbook

Africa is not a constrained version of the Valley playbook. It is a different game with different winning moves.

5. Possible directions (Africa-first)

  1. WhatsApp agent runtime for SACCOs/SMEs — the wedge: every SACCO already runs on WhatsApp groups; give them member-service agents + organized data behind it. Customer pays from day one.
  2. Vertical agritech data flywheel — capture farm-level data through agents, organize it, sell the intelligence back. Agriculture is the named next frontier.
  3. African-language eval & data studio — the Swahili/Sheng gap is a business: eval datasets, fine-tuning data, and testing harnesses for anyone shipping AI in East Africa.
  4. Offline clinic/school agents — on-device small models where connectivity can't be assumed; EduManage's intelligent-school-ops thesis extended to health.
  5. M-Pesa/KRA MCP marketplace — every Kenyan business integration (payments, tax, SMS) packaged as agent-usable tools. Boring, lucrative, defensible.
  6. Enterprise agent reliability layer — evals, guardrails, human-in-the-loop approvals, audit trails as a product for companies deploying agents (Orosz's "verification is the scarce resource," productized).
  7. USSD/SMS-fallback gateway — one API that lets any agent reach users with no smartphone and no data. The unsexy infrastructure nobody builds and everybody needs.

6. The integration roadmap (sequenced)

7. Open questions