AI-Native Code Generation

Context, agents and systems thinking — a talk at Nokia, Bangalore, for an audience of ~150 engineers.

September 12, 2026 · Talk + 60 slides

Amarnath presenting the talk outline, “The talk, in five beats”, on the big screen at Nokia — Opening the talk · Photo: Srushti Kadavadi

I gave a talk on AI Native Software Engineering (slides below) at Nokia's Bangalore office on Sept. 12. The talk was organised by Deep Tech Stars, and had an audience of ~150. This was my first ever time speaking in front of so many people. The audience comprised of software engineers and tech leaders from across the industry.

Amarnath on stage in front of the audience at Nokia's Bangalore office — On stage

Amarnath presenting to a full room of attendees — On stage

Amarnath beside the “Performance = intelligence × context” slide — Performance = intelligence × context · Photo: Boomika Shivkumar

It was great to speak to so many curious people after the talk. What was supposed to be a 45 min presentation, extended into a 75min conversation.

Amarnath in front of the Nokia sign outside the office — After the talk

Amarnath with Boomika Shivkumar in front of the Nokia sign — After the talk

Thanks to my co-speakers Asha Rajput and Malla Avinash Kumar Reddy for their informational talks. Thanks to the attendees for listening to us!

Thanks Nihal Kashinath for the opportunity.

Group photo of attendees and speakers in the event hall — Photo: Amarnath Prasad

Originally posted on LinkedIn. Photos shared by attendees are credited to them.

The slides

All 60 slides, as presented. Open the deck to view it full screen. Use ← → to move between slides, and press Esc or Exit presentation to come back to the page.

AI-native development

AI-Native Code Generation

Amarnath Prasad

Founder of Cralo AI Unacademy · JioHotstar · Memfold AI

AI-native development

AI-Native Code Generation

Amarnath Prasad

Founder of Cralo AI Unacademy · JioHotstar · Memfold AI

AI-native developmentThe new engineer

You hire a brilliant engineer

Now imagine their first day.

AI-native developmentThe new engineer

No access. No feedback.

No repository access

No CI. No logs. No shell.

AI-native developmentThe new engineer

No context. No verification.

No docs, JIRA, or Slack history

No way to run the tests

AI-native developmentThe new engineer

You do the handoff over calls

That’s how most teams use AI agents today.

AI-native developmentThe thesis

Prompting isn’t enough. Engineer the system.

AI-native developmentThe roadmap

The talk, in five beats

01
What AI-native development actually is
02
Prompt → context → loop engineering
03
A live walkthrough of an AI-native product
04
Large codebases, then software factories
05
Five principles to take home

AI-native development

The familiar lifecycle

AI-native developmentThe familiar lifecycle

The software development lifecycle

PRD

Architecture

Modules

Tasks

Code

Review

QA

Ship

Monitor

AI-native development

The AI-native lifecycle

AI-native developmentThe AI-native lifecycle

Same flow. New execution.

PRD

Agent drafts spec.md. You correct and approve.

Architecture

Agent suggests an approach. You decide.

AI-native developmentThe AI-native lifecycle

Plan once. Build in parallel.

Tasks

Agent writes plan.md.

Build

Agents in git worktrees. One branch per task.

AI-native developmentThe AI-native lifecycle

Verify first. Review the PR.

Review

Agents verify first. You review the PR.

Ship

Automated. Agents watch the signals.

AI-native developmentOwnership

What stays human

Humans

Stack. Architecture. Product. Merge. Release.

Agents

Draft. Build. Verify. Iterate until checks pass.

Humans: decisions and gates. Agents: execution.

AI-native developmentThree eras

From prompts to autonomous loops

2022–23

Prompt

Craft the ask

2024–25

Context

Give it the knowledge

2026–

Loop

Act, verify, learn

Context Engineering

Context engineering

Give the agent the right information.

Context Engineering

Performance = intelligence × context

Right context, right time — not a dump.

Context Engineering

Connect everything

Docs

JIRA

Slack

Incidents

Design

Code

AgentSearch → retrieve

Retrieve just-in-time. Connected ≠ pre-loaded.

Context Engineering

Let the agent retrieve it

jira issue view PROJ-123
 
rg "invoice" packages/billing

Search at runtime. Pull only what the task needs.

Context EngineeringConnecting sources

MCP: structured tool servers

01
Register in Cursor / Claude Code
02
Jira, Slack, docs as typed tools
03
Schemas the agent discovers

Context Engineering

Keep context in the repo

Context EngineeringKeep context in the repo

AGENTS.md

# how this repo works
pnpm test --filter payments
 
# never
# edit generated files in gen/
# raw SQL — use make db-migrate
 
# deploy
main → prod · feature/* → preview

Always loaded. Keep it small.

Context EngineeringSkills

Skills load on demand

AGENTS.md is always on. Skills are not.

01
Point to the source of truth
Don’t duplicate docs.
02
Be opinionated about workflows
03
Security rules stay always-loaded

Context EngineeringKeep context in the repo

Keep context in the repo

my-repo/

.skills/On demand

deploy/How we ship

AGENTS.mdShared rules

packages/

payments/

AGENTS.mdPayments rules

billing/

AGENTS.mdBilling rules

Small root file. Skills on demand. Package-specific detail.

Loop Engineering

Loop engineering

Give the agent a way to act, check, and improve.

Loop Engineering

The whole idea, in one line

while :; do
  cat PROMPT.md | claude-code
done

Fresh context each iteration. Progress lives in the repo.

Loop Engineering

Run until the work is done

Re-read spec

Work one task

Check tasks.md

Observe · repeat

Exit when fully checked off. Failing checks keep the loop alive.

Loop Engineering

Close the feedback loop

CI fails

Read logs

Hypothesize

Fix

Re-run

A failed check becomes the next action.

Loop Engineering

Verification

Loop EngineeringVerification

Verification is the exit condition

Test suite + CI

Exit code 0

Playwright

Screenshot → compare → fix

Staging smoke

Real deploy, end to end

Loop EngineeringVerification

The verification pyramid

Human judgment

Adversarial review

Skills · lint rules

Tests · types · linters

Loop Engineering

Runtime

Loop EngineeringRuntime

Remote Docker container

Isolated. Reproducible. Disposable.

Loop EngineeringRuntime

Give the run only what it needs

01
Scoped tokens from your vault
02
Only the connectors this task needs
03
Staging-only deploy credentials

Jira · Slack · GitHub · database

Loop EngineeringRuntime

One disposable run

docker run --rm \
  -e JIRA_API_TOKEN -e GITHUB_TOKEN \
  agent-loop:latest

Inject scoped credentials. Keep secrets out of logs.

Loop EngineeringRuntime

Set boundaries before the run

01
Restrict commands, paths, and network access
02
Set time, retry, and cost limits
03
Stop and escalate when progress stalls
04
Tag agent PRs. Keep human merge approval.

Loop Engineering

Event triggers

Loop EngineeringEvent triggers

Machines initiate. Humans approve.

Event

GitHub · Jira Sentry · Slack

Agent loop

Plan · build · verify

Human at merge

Reviews the PR

Loop Engineering

Make the next loop better

Loop EngineeringCompound knowledge

Teach the system, not the session

Ship a task

Capture the lesson

Encode it

Save the lesson in AGENTS.md or a skill.

Loop EngineeringCompound knowledge

Turn lessons into checks

01
eslint-plugin / no-raw-sql
02
Pre-commit hook
03
Codemod
04
CI gate

Repeated mistakes become repeatable checks.

Loop Engineering

Design loops that prompt your agents

“You shouldn’t be prompting coding agents anymore.” — Peter Steinberger, creator of OpenClaw

A live example

Large-codebase playbook

The same loop needs a repo it can navigate.

Large-codebase playbook

Make the repo answer questions

01
Fast whole-repo search
Grep at this scale is infrastructure.
02
Make the repo structured
Clear modules. Obvious boundaries.
03
Rules beside the code
Local AGENTS.md, not a dump.

Navigability is your job. Retrieval is the agent’s.

Large-codebase playbook

Keep the codebase maintainable

01
Characterization tests before refactoring
02
Human gates on architecture and merge

Agents fail silently on cross-cutting change.

Large-codebase playbook

Make “done” mean something

01
Tests + perf budgets in the definition of done
02
Target 2–3× faster, safely — not 10× without review

Time to an accepted change — not time to generate code.

Software factories

Extend the loop until the operating process is the product.

Software factoriesIndustry adoption

Most teams: inline AI + agent mode

01
Accept / reject every line
02
Review plans and diffs
03
A human still sits in every loop

Software factoriesIndustry adoption

The frontier: triggered, autonomous

70%+

PRs involve agents

Uber

5–30h

With nobody watching

Anthropic

Claude Code and the CLIs

Software factories

Extend the loop → the factory

Signals → triage

Plan → build

Verify → ship

Observe · repeat

Monitor → production feedback → back into the queue

Take home

Five principles to carry into Monday.

Take homeFive principles

Don’t assume capabilities

Your agent is ultra capable. Don’t underestimate it.

Take homeFive principles

Give as much autonomy as possible

Don’t sit in every loop.

Take homeFive principles

Guardrailed access to real tools

Logs, CI, shell — in a sandbox.

Take homeFive principles

Connect everything. Retrieve just-in-time.

Right context, right time.

Take homeFive principles

Let the agent test its own work

Run tests. Use the browser. Check it the way a human would.

Take homeFive principles

Five principles to take home

01
Don’t underestimate the agent
02
Give it autonomy
03
Provide guardrailed access to tools
04
Connect context. Retrieve when needed.
05
Let it test its own work

Take homeKeep up

Be in the room where it happens

Follow the practitioners and the labs

Practitioners

@dexhorthy

@hraness

@addyosmani

@steipete

Labs & products

@AnthropicAI

@OpenAI

@cursor_ai

Take homeThank you

Questions?

Engineer the system

@amrnth0 amrnth.com