Indonesia 2026
Ahmad Rosid
Co-Founder & CTO at Helipod.io
01 / 25
About me
Ahmad Rosid
Based in
Metro City
Lampung, Indonesia
Start
Self-taught
SMK graduate · no CS degree
~10y
Software engineering
Building products full-time
8y
Remote working
US & international teams
2025
Orchids · YC W25
Building a coding harness for AI
2026
Figma acquisition
Role not in the deal · laid off
Now
Helipod.io
Co-founder & CTO
Goal of this talk
Get the job done.
Don't burn out.
Done
Scoped change. Green tests. Reviewable PR.
Not burnout
No babysitting. No re-explaining. No midnight thrash.
Built at Orchids
#1 on public benches.
App Bench
#1 · 76%
UI Bench
#1 · 30.08
What is Harness?
The harness is what makes an LLM useful. On its own, a model is just next-token prediction — really good at producing text.
LLM alone
Predicts the next token. Writes text. Can't touch your repo.
+ Harness
Gives the model hands — so it can actually work in your codebase.
01Read files
02Write files
03Edit files
04Run bash commands
Hands get it into the repo. They don't make it succeed.
Tools only
Hands aren't enough
-
01
No map → wrong place
-
02
No playbook → random steps
-
03
No checks → "done" is a guess
-
04
You babysit until midnight
Map + skills + checks
Design the workplace
-
01
Map → knows where
-
02
Skills → follows your flow
-
03
Guardrails → knows wrong
-
04
You review & ship
What doesn't fix it
Smarter model ≠ better harness
"Just buy a smarter model." Fable 5 is 2× Opus — and still blind without a harness.
Haiku 4.5
$1 / $5
Sonnet 4.6
$3 / $15
Opus 4.8
$5 / $25
Per 1M tok
Fable 5
$10 / $50
What is Harness Engineering?
Harness engineering is designing an environment for the agent — so it has a better success rate when working on a task.
Environment for your agent
01
Map the codebase
AGENTS.md, ARCHITECTURE.md — agent knows where to look.
02
Skills & workflow
Skills — plan, debug, ship. Agent follows your workflow.
03
Mechanical guardrails
Tests, lint, CI — the agent knows when it's wrong.
04
Feedback loop
When it fails, don't try harder — find what's missing, have the agent commit the fix.
Harness engineering · 01
Map the codebase
A map tells the agent where to look, so it doesn't load the whole repo. Keep context manageable.
/
AGENTS.md
Navigate the repo and verify with tests.
/
ARCHITECTURE.md
Database, language, and engineering decisions.
Harness engineering · 02
Skills & workflow
Skills are playbooks. The agent follows your process instead of inventing one each time.
/plan
Plan
Scope the change before any edit.
/debug
Debug
Reproduce, isolate, then fix.
/ship
Ship
Tests pass. PR ready.
Harness engineering · 03
Mechanical guardrails
The agent needs a way to know when it's wrong. Wire that into the loop.
01
Tests
Run them. Fail means stop.
02
Lint
Style and safety before merge.
03
CI
Same checks every push.
Harness engineering · 04
Feedback loop
When the agent fails, the fix is never "try harder." Find what's missing and put it in the repo.
01
It fails
Stuck or wrong. Don't re-prompt harder.
02
Find the gap
Missing tool, guardrail, or doc.
03
Commit the fix
Have the agent write it. The harness gets stronger.
Live example
See this technique in action
A real harness you can clone. AGENTS.md, ARCHITECTURE.md, agent teams.
GitHub github.com/ahmadrosid/nakamaAGENTS.md
Map for the agent.
ARCHITECTURE.md
Boundaries and data flow.
Open source
Clone it. Read the harness.
The gist
Rule of thumb
Your job: scope, map, and rules. The agent's job: execution.
-
01
You decide scope
-
02
You write the map
-
03
You define "done"
-
04
Agent edits — you review
-
05
Fix the harness, not the model
-
06
Speed over perfection
Bonus · Compound Engineering · 01
Compound Engineering
Decide what to build, plan how to build it, then work the plan — with decisions written down in the repo.
01
Brainstorm
What to build. Scope and requirements.
02
Plan
How to build it. Units and files.
03
Work
Execute the plan. Verify as you go.
Bonus · Compound Engineering · 02
Durable docs
Decisions live in the repo, not just in chat. Each doc carries the record forward.
docs/brainstorms/
requirements
What to build. Scope, actors, acceptance examples.
docs/plans/
implementation plan
How to build it. Units, files, test scenarios.
Bonus · Compound Engineering · 03
Why the record helps
The agent reads the doc — not your last chat — so work picks up where you left off.
01
Survives context loss
New session. Same decisions.
02
Agent can resume
Scan units. Continue the plan.
03
Decisions inspectable
You can review what was decided and why.
Bonus · Token optimization · 01
Spend context wisely
Agents burn tokens on noisy command output and repeated file reads. Shrink both sides of the loop.
01
Compress output
Filter git, test, and lint noise before it hits context.
02
Query structure
Ask the graph — don't grep the whole repo again.
03
Same harness
Both plug into Cursor and Claude Code.
Bonus · Token optimization · 02
RTK — Rust Token Killer
CLI proxy that filters and compresses command output. Single Rust binary. Hooks rewrite bash before the agent sees it.
GitHub github.com/rtk-ai/rtk60–90% savings
git, tests, lint, docker — compact by default.
rtk init -g
Auto-rewrite hook for Cursor & Claude Code.
rtk gain
Track token savings per session.
Bonus · Token optimization · 03
codebase-memory-mcp
Indexes your repo into a persistent knowledge graph. One structural query replaces dozens of grep/read cycles.
GitHub github.com/DeusData/codebase-memory-mcpindex_repository
Build the graph once. Query in milliseconds.
trace_path
Who calls this? What does it call?
~99% fewer tokens
vs. file-by-file exploration.
Bonus · UI iteration · 01
Iterate UI fast
Four mockups in ChatGPT beat one slow coding pass. Pick visually, then implement.
- 01 Attach a screenshot — ask for 4 UI variants
- 02 Compare layouts — pick a direction
- 03 Give the winner to your agent to build
Copy-ready prompt
Generate image: of 4 different UI designs from the attached image to enhance its UI/UX, and label them A through D.
Bonus · UI iteration · 02
Quick edit in design mode
Select an element in Trae. Describe the change — the agent edits in place.
Bonus · Productivity · 01
Dictate, don't type
You type all day with AI. When your hands are done, speak your prompts instead.
Free & open source
Wispr Flow
Paid & polished
Thank you
Build the harness.
Write the record.
Steer the work. Let agents execute. Make the next session smarter.
Questions? · github.com/ahmadrosid