AI-Native Infrastructure Research

Chapter 01

Three lenses. One discipline.

Gergely Orosz, Petar Ivanov and ByteByteGo arrive from different angles—organization, engineering discipline and system architecture—but converge on verification as the scarce resource.

ORGANIZATION

Gergely Orosz

Slow down to speed up. AI increases individual output faster than teams can absorb it. Review by blast radius, keep work units small, and optimize CI/CD for risk reduction—not merely speed. source

ENGINEERING

Petar Ivanov

An agent is mostly not a model. Routing, tools, schemas, retries, state and formatting are deterministic machinery. Unit-test that machinery; judge only the behavioral surface; keep production evals running. source

ARCHITECTURE

ByteByteGo

Choose the cheapest pattern that works. RAG is predictable retrieval; agents reason and act; MCP standardizes tool access. Read and write paths need different architectures. source

“Production AI is systems engineering, not prompting.”

Research synthesis across all three writers

01

Humans stay in the loop for high-blast-radius actions.

02

Generation and verification must be separate concerns.

03

Reliability lives in the deterministic substrate.

04

Production itself is the eval that never ends.

Chapter 02

The company OS for an agentic business.

Do not build a monolithic “mother agent.” Build a layered operating system where intelligence is replaceable and every consequential action crosses a policy boundary.

01 · Human edge

WhatsAppPrimary customer and staff interface

SMS / USSDGraceful fallback under weak connectivity

Role-specific assistantsSales, operations, finance, support

02 · Control plane

Agent orchestratorPlans, selects tools and delegates

Durable workflow runtimeCheckpoint, resume, retry one step

Policy gatewayAllowlist, schema validation, approvals, audit

Evaluation planeUnit tests, scenario judges, production alerts

Model routerCost, capability and fallback policy

03 · Business substrate

MCP toolsM-Pesa, KRA eTIMS, inventory, tickets

Operational databaseSource of truth outside the model

Vector memoryDocs, tickets and run history

Queue + trace storeAsync work, replay evidence, cost attribution

The boundary that matters

The model proposes; the machine disposes. Models can decide which approved capability to request. Deterministic code validates parameters, enforces permissions, writes state, checks idempotency, pauses for a human and records what happened. The research’s “Company OS” pattern recommends Temporal or LangGraph, MCP tools, Postgres/pgvector or Qdrant, LiteLLM, an eval layer and WhatsApp as the interface. The synthesis is explicitly not a single canonical source.

Draw the model / machine boundary

Test deterministic code normally. Reserve probabilistic judges for behavior.

Ship all three eval layers

Unit tests → LLM-judge scenarios → continuous production evals with alerts.

Use RAG before autonomy

If retrieval plus generation solves it, avoid the extra cost and failure surface of a loop.

Split read paths from write paths

Different risks require different tools, permissions and review bars.

Review by blast radius

Fast-path low-risk output; inspect plans, tests and schemas for consequential work.

Put a gateway before every tool

Inject credentials at runtime; validate parameters; enforce allowlists and approvals.

Design for recovery

Checkpoint after each step and make writes idempotent so restarts resume instead of replay.

Trace the trajectory

Record calls, arguments, latency, tokens, retries, guardrails, approvals and cost—not just the final answer.

Spend judgment, not tokens

Ship fewer things that matter. “More AI” is a cost, not a virtue.

Sell verification

The wedge is trust in agent output: evals, guardrails, audit and cost attribution.

Chapter 03

Choose a pattern before a model.

Architecture follows the work. Select the task shape below to see the smallest reliable pattern supported by the research.

Use RAG.

One retrieval plus one generation is cheaper, predictable and easier to debug. Keep facts in a source you can update and cite.

ByteByteGo · RAGs vs Agents

Low blast radius

Read

Search, summarize, draft. Log the evidence and automate freely within bounded scope.

Medium blast radius

Prepare

Generate a proposal or change set, then put a human at the commit boundary.

High blast radius

Write

Payments, bulk changes and destructive actions require deterministic policy and explicit approval.

Chapter 04

Africa’s constraints are product primitives.

The opportunity is not to localize a Western agent after launch. It is to make connectivity, language, payments, trust and regulation first-class in the architecture.

~⅓of Africans have regular internet access, per the report’s World Bank WDR 2026 citation

~70%use entry-level smartphones, per the same research summary

20+point open-model performance drop on African languages vs English in AfroBench

<2%of H1 2026 African funding went to genuinely AI-native companies, per cited reporting

DIRECTION 01

WhatsApp-native runtime

Give SMEs and SACCOs an agent where business already happens, not another dashboard.

Customer: SACCOs, banks, e-commerce, clinics
Wedge: No app install; familiar chat interface

DIRECTION 02

Vertical data flywheel

Capture one hard-to-recreate dataset and turn each deployment into better intelligence.

Customer: Agribusiness, insurers, counties
Wedge: Crop imagery + Swahili agronomy workflows

DIRECTION 03

African-language eval studio

Apply Snorkel’s Data 2.0 model to Swahili, Hausa, Yoruba and Amharic.

Customer: Global labs, telcos, governments
Wedge: Expert tasks, rubrics and RL environments

DIRECTION 04

Offline-first agent appliance

Small models at the edge, queued work and sync when connectivity returns.

Customer: Clinics, schools, NGOs
Wedge: Reliable work with zero connectivity

DIRECTION 05

Local MCP marketplace

Expose fragmented local rails once so every agent builder can reuse them.

Customer: East African agent builders
Wedge: M-Pesa, KRA eTIMS, bank and ERP connectors

DIRECTION 06

Reliability & eval layer

Package durability, evaluation, audit and cost attribution for early enterprise adopters.

Customer: Banks, SACCOs, insurers
Wedge: Sell verification—not another agent

DIRECTION 07

USSD / SMS fallback gateway

When data drops, the workflow continues through feature-phone-compatible rails. This is not graceful degradation as an afterthought; it is a first-tier interface for markets global frameworks ignore.

Customer: Telcos, agri-input companies, micro-lenders
Wedge: Continuity across device and network constraints

Regulatory watch

Kenya’s Data Protection Act and the ODPC’s July 2026 draft AI guidance bring DPIA triggers, 72-hour breach notification and cross-border transfer rules into the design. The proposed Kenya AI Bill 2026 may add audit-trail obligations; KICTANet’s review warns that its current drafting could make dominant API-dependent development technically noncompliant. Track the bill rather than treating it as settled law. KICTANet review

Chapter 05

What the breakouts actually sold.

The signal is not “AI is valuable.” It is that value accrues to scarce data, durable execution, enterprise data gravity and developer defaults.

Company	Proof point	The wedge	Transferable lesson
Scale AI	$14.3B Meta investment for 49% at about $29B, June 2025	Neutral data foundry for frontier labs	Neutrality was the product—and the strategic deal damaged it. source
Snorkel AI	$350M Series E at $3.5B, reported 22 Sep 2026	Data products: expert tasks, environments and rubrics	Move from labeling hours to engineered data products. Reuters
Databricks	$5.4B ARR run rate, Jan 2026; 65% YoY growth	Enterprise data gravity	AI workloads follow the governed operational data. source
Temporal	$12.55B reported valuation, Sep 2026; $250M ARR run rate	Durable execution for work that must survive failure	Recovery separates production systems from demos. source
LangChain	$1.25B valuation, Oct 2025	Open-source developer default → agent engineering platform	Developer mindshare can become observability and eval demand. TechCrunch
InstaDeep	~$550M completed acquisition; ~$682M announced incl. milestones	Decision AI and research talent built in Tunis	Africa can export world-class vertical intelligence; acquisition figure varies by source.
Aerobotics	$26M+ raised; used in 18 countries	Drone imagery → crop insight → more imagery	A vertical, local data flywheel can compound globally.

The flywheel

Capture → label and enrich → train → deploy → capture harder edge cases. The research argues this is won vertically. In Africa, that can mean crop-health imagery, dialect speech, factory process data or regulated business workflows—not a horizontal labeling marketplace.

Chapter 06

Use each platform for its sharpest edge.

Do not choose a single “AI provider.” Separate interface, model, orchestration and tool standards so each can evolve independently.

Reasoning + skills

Anthropic

Claude API for capable reasoning
MCP as the shared tool protocol
Agent Skills for repeatable workflows
Prompt caching and Batch API for cost control

Haiku 4.5$1 / $5

Sonnet 5$2 / $10

Opus 5.5$4 / $20

Reported Sep 2026 prices per 1M input / output tokens. Production access requires paid billing.

Low-cost multimodal + hosting

Google

AI Studio for prototyping
Vertex AI for IAM, observability and enterprise controls
ADK with A2A support
Agent Engine and managed MCP services

Gemini 2.5 Flash~$0.30 / $2.50

Gemini 3.6 Flash$0.75 / $3.75

3.6 Flash pricing reported as introductory through 31 Dec 2026; verify before budgeting.

Distribution

Monthly run-cost sketch

Model Monthly conversations Input tokens / conversation Output tokens / conversation WhatsApp replies / conversation

$0.00

estimated monthly model + WhatsApp message cost

Model$0.00

WhatsApp$0.00 · KSh 0

Directional only. Uses the static prices above, assumes every reply is charged at the reported Kenya service-message rate, and excludes caching, gateways, hosting, storage, taxes and retries. Verify provider terms before a commercial decision.

Routing policy

Start with OpenRouter for low-ops exploration, then use self-hosted LiteLLM for production routing and data control. Pin multi-turn sessions to one provider to preserve cache reuse; only fail over to models that pass the same tool-calling evals. The research estimates classifier-based gateways may add 50–150ms to time-to-first-token.

Chapter 07

Use Muse as the founder’s control room.

Treat Muse less like a chat window and more like an operating loop: durable context, reusable skills, bounded delegation, scheduled checks and reviewable artifacts.

01

Define the outcome

One business result, constraints, authority boundary and evidence of done.

02

Capture context

Keep stable principles separate from changing project state and source material.

03

Encode the skill

Turn repeated work into a named playbook with scripts, references and checks.

04

Delegate a run

Give the agent a bounded task; keep high-risk writes behind approval.

05

Verify evidence

Review output, logs, tests and the claimed state—not the agent’s confidence.

06

Improve the system

Convert failures into tests, stronger tools, sharper context and new guardrails.

While you are present

Interactive design

Use critique, prototypes and decision artifacts to shape the system. Ask for options only where a choice changes the product.

While you are away

Bounded monitoring

Schedule read-only research, reconciliation, test runs and status checks with explicit delivery rules.

When action is needed

Human checkpoint

Pause before sending, paying, publishing, deleting or changing production. Resume from the saved state after approval.

The unattended-agent contract

Every background agent should have: a stable run ID; a step budget, token budget and wall-clock limit; checkpointed state; idempotency keys for writes; one owner for retries; a cross-provider fallback that passed evals; an approval boundary; and a final evidence packet. This turns “work while I’m off” from an open-ended promise into a resumable, inspectable process.

Agent build status: specified, not deployed

This artifact is a static research and operating brief; it cannot run a persistent background worker. No unattended production agent, scheduler or external account connection was created here.

Deployment handoff

Choose one bounded first job: read-only campaign reconciliation, research monitoring or test execution—not sending, paying or publishing.
Provision the runtime: a server-backed Muse automation or durable workflow worker with a stable run ID, checkpoint store and queue.
Connect only the required services: the owner must authorize the target account through its approved connection flow. For a WhatsApp-facing product, that means an owned Meta Business/WhatsApp Cloud API setup; for model calls, a paid provider account or routed gateway. Never paste raw credentials into a prompt or this page.
Install the control plane: typed tools, read/write separation, idempotency keys, allowlists, run budgets, approval pauses and immutable audit records.
Ship the eval pack: deterministic unit tests, representative scenario tests and production alerts with explicit failure thresholds.
Run one supervised canary: verify the evidence packet and recovery behavior before enabling a schedule.

Minimum operating manifest

job → allowed tools → data scope → run/step/token/time budgets → retry owner → checkpoint key → approval boundary → delivery target → evidence packet

Access needed for deployment: approval to create a persistent server-backed worker, plus connected access only to the specific channel and data source chosen for the first job. None is needed to use this research brief.

Chapter 08

A 90-day direction for Ansai.

A capital-constrained sequence: prove one workflow, harden the control plane, then package the local infrastructure that repeats.

Days 1–30

One workflow, end to end

WhatsApp Cloud API test number
OpenRouter + Gemini 2.5 Flash comparison
LiteLLM routing from day one
Read-only pilot with real evaluation cases
Instrument cost, latency and failures

Days 31–60

Make it recoverable

Add durable execution
Checkpoint every step
Separate read and write tools
Require schemas and idempotency
Ship unit, scenario and production evals

Days 61–90

Package the repeatable layer

Expose business tools as MCP servers
Turn workflows into skills
Add an approval console and audit trail
Test SMS / USSD fallback where required
Sell the reliability layer with measured ROI

Defer deliberately

Do not train a foundation model, compete with horizontal data clouds, bet the company on x402 payments, depend on Cloudflare Pay Per Crawl, or build multi-region self-hosted inference before a use case demands it. The protocol layer is real; the autonomous agent economy is not yet proven. The research notes that only 0.6–7.5% of x402 volume may be genuinely agentic and roughly half of activity may be gamed.

Appendix

Evidence, provenance and open questions.

The report was assembled from web-index research on 24 September 2026. Only The T-Shaped Dev homepage was live-verified; several major claims rely on reputable secondary reporting or search-index extracts.

Gergely Orosz · code reviewnewsletter.pragmaticengineer.com

Petar Ivanov · agent testingpetarivanov.tech

ByteByteGo · MCP vs RAG vs agentsblog.bytebytego.com

ByteByteGo · RAGs vs agentsblog.bytebytego.com

Grab agent architectureblog.bytebytego.com

Agent state architecturedev.to

Snorkel AI fundingReuters

LangChain fundingTechCrunch

Temporal fundingValue Add Ventures

African AI fundingEcofin Agency

Africa AI landscape 2026Africa Global Forum

Kenya AI Bill reviewKICTANet

Cloudflare pay-per-use shiftPPC Land

x402 hype checkAlphaWire

LLM gateways comparisonOpper.ai

WhatsApp pricing analysisLinkedIn analysis

Carry these caveats forward

Orosz’s Anthropic and OpenAI dispatches were paywalled; the research used secondary GitHub notes. Treat the 70–90% Claude-generated-code claim as reported, not independently verified.
Snowflake market-cap sources conflict. InstaDeep’s acquisition is reported as about $550M completed versus about $682M announced including milestones.
Recent Labelbox, Weaviate and Inngest funding; pure-play MCP startup funding; and Meta Model API details were not verified in this run.
Direct M-Pesa support for provider API billing was not verified. Provider access and billing details can change.
Cloudflare Pay Per Crawl remained private beta in the research; agent-payment standards were forming, but measured usage was largely non-agentic.