TRIPOS·BUILT BY AGENTS

§ Behind the scenes · How it's made

One human.
Eleven agents.
Two sentinels who never sleep.

TripOS was designed, built, deployed, and is monitored by a team of AI agents working for one founder. There's no engineering team. There's no on-call rotation. There's no Jira. The agents do the work — the human ships the vision.

✈Meet the teamSee the loop ↓

Krish Mehta · operator · the only human in the loop

§ Live demo · Skip the prose, watch them work

Tap the control plane.
Watch SEV-2 close itself.

A synthetic incident drill — what happens when a real one fires at 3am. Each tier owns its lane. The log on the right is what the agents say to each other. ~7 seconds end to end, no human in the loop.

Incident response · interactive demo
01 · Control plane
02 · First respondersEach owns a remediation lane

WARDEN

First responder · auth & security

standby

ATLAS

Reroute · architecture

standby

FORGE

Rollback · backend

standby

PILOT

Rotate · devops

standby
03 · Quality gateHolds the all-clear

CHECK

Verify · QA probes

standby
04 · User commsCloses the loop with anyone affected

CONCIERGE

Remediate · runtime AI

standby
Also on fleet · not pagedIdle by design
PIXEL
VAULT
DRAFT
VERSE
REACH
Agent feed · idle
  • PIXEL · committed to feature/live-mode-pill
  • CHECK · type-check passed · 3.1s · 0 errors
  • FORGE · /api/trips/[id]/replan · +12 / −4
  • PILOT · preview deploy READY
  • REACH · sitemap regenerated · 36 routes
Tap CONTROL above to fire a SEV-2 drill

§ The thesis · Why this exists

SaaS used to need
ten people to ship.
It needs one now.

The hard part of building a software company was never the typing. It was coordinating ten specialists — frontend, backend, devops, design, copy, QA, security — to all move in the same direction without dropping context.

That coordination problem is what AI agents solve. Give each one a charter, a shared codebase, and a human who can taste-test the output, and you have an engineering org. TripOS is the proof. Eleven agents put it on the internet. Two more keep it from breaking. Krish points the ship.

This is what the next decade of SaaS looks like. We figured we'd build one and show you.

§ The team · 11 builders

Meet the engineering
organisation.

Each agent owns a charter. They share the codebase, share context, and pass work between each other through pull requests and shared memory. They argue when they disagree. The human breaks ties.

ATLAS

System architect

opus

Owns the shape of the thing. Schemas, API contracts, data flow.

Runs on
Opus 4.7
Context
1M tokens
Input rate
$15.00 / M
Output rate
$75.00 / M

Holds the whole codebase in head when planning cross-cutting work.

Tools

ReadGrepGlobTaskCreateSkill

Hands off to

PIXELFORGEVAULTWARDEN

Shipped

  • 55 data models
  • 127 API routes
  • Trip generation pipeline

PIXEL

Frontend craftsman

sonnet

Components, layouts, animations. Lives in TSX.

Runs on
Sonnet 4.6
Context
200k tokens
Input rate
$3.00 / M
Output rate
$15.00 / M

Fast iteration loops. UI work needs many small passes, not deep planning.

Tools

ReadEditWriteGlobBash

Hands off to

DRAFTCHECK

Shipped

  • 104 React components
  • Live Mode mobile-first
  • Custom primitives

FORGE

Backend engineer

opus

API routes, server actions, business logic. Where the verbs live.

Runs on
Opus 4.7
Context
1M tokens
Input rate
$15.00 / M
Output rate
$75.00 / M

Business logic spans dozens of files; reasoning across them needs the bigger model.

Tools

ReadEditWriteBashGrep

Hands off to

VAULTCHECKWARDEN

Shipped

  • Replan engine
  • Geofence checkin
  • Booking sync

VAULT

Database keeper

sonnet

Schema, migrations, query plans, indexes that matter.

Runs on
Sonnet 4.6
Context
200k tokens
Input rate
$3.00 / M
Output rate
$15.00 / M

Schema work is local and pattern-driven. Sonnet is the right cost/precision trade.

Tools

ReadEditBash

Hands off to

CHECKPILOT

Shipped

  • Zero-downtime migrations
  • Composite indexes
  • Soft deletes

WARDEN

Security & auth

opus

Auth, CSP, rate limits, secret rotation, RBAC. Boring on purpose.

Runs on
Opus 4.7
Context
1M tokens
Input rate
$15.00 / M
Output rate
$75.00 / M

Security mistakes are expensive. Reasoning capacity > token cost here.

Tools

ReadGrepEditBashWebFetch

Hands off to

CHECKATLAS

Shipped

  • Owner-gated brand
  • CSP allowlist
  • IP rate limits

PILOT

DevOps

haiku

Deploys, env vars, edge functions, preview links.

Runs on
Haiku 4.5
Context
200k tokens
Input rate
$0.25 / M
Output rate
$1.25 / M

Deploy ops are scriptable and repetitive. Haiku is fast and cheap for this.

Tools

BashReadWebFetch

Hands off to

SENTINEL-1SENTINEL-2

Shipped

  • Preview-per-PR
  • Env-pull on every clone
  • Error tracking wired

DRAFT

Design system

sonnet

Departure Board tokens, type ramp, brand consistency police.

Runs on
Sonnet 4.6
Context
200k tokens
Input rate
$3.00 / M
Output rate
$15.00 / M

Visual judgment + token-level edits. Sonnet hits both well.

Tools

ReadEditWriteGlob

Hands off to

PIXELVERSE

Shipped

  • Sand/ink/cobalt/coral palette
  • Display + mono pairing
  • 80 brand assets

CHECK

Quality control

sonnet

TypeScript correctness, integration verification, regression hunting.

Runs on
Sonnet 4.6
Context
200k tokens
Input rate
$3.00 / M
Output rate
$15.00 / M

High-volume review pass over diffs. Sonnet absorbs the throughput.

Tools

BashReadGrep

Hands off to

PILOTATLAS

Shipped

  • Zero TS errors at HEAD
  • Smoke flows on critical paths

VERSE

Copywriter

haiku

Microcopy, error messages, marketing pages. Words that don't waste seconds.

Runs on
Haiku 4.5
Context
200k tokens
Input rate
$0.25 / M
Output rate
$1.25 / M

Short generative bursts on bounded scope. Haiku ships words at speed.

Tools

ReadEditWrite

Hands off to

DRAFTPIXEL

Shipped

  • /agencies B2B page
  • Onboarding copy
  • Empty-state voice

REACH

SEO & growth

haiku

Meta tags, sitemap, structured data, analytics events.

Runs on
Haiku 4.5
Context
200k tokens
Input rate
$0.25 / M
Output rate
$1.25 / M

Structured-output, repetitive work. Haiku is the cost-right answer.

Tools

ReadEditWebFetch

Hands off to

DRAFTPIXEL

Shipped

  • Per-page metadata
  • Sitemap generation
  • Funnel events

CONCIERGE

Runtime AI

sonnet

The agent that ships *to users*. Replan, alerts, geofence reasoning.

Runs on
Sonnet 4.6
Context
200k tokens
Input rate
$3.00 / M
Output rate
$15.00 / M

User-facing latency budget. Sonnet hits sub-second responses on small inputs.

Tools

WebFetchSkill

Hands off to

End user

Shipped

  • Live Mode reroutes
  • Weather-aware swaps
  • Closure detection

§ Stack & cost · Which model runs each charter

Three model tiers.
One bill at the end.

Each agent runs on the model tier that fits its charter — cheap and fast where the work is repetitive, deeper reasoning where the work spans the repo. Public per-token pricing, no markup, no per-seat fees.

Opus 4.7

opus
Context
1M tokens
Input
$15.00 / M tok
Output
$75.00 / M tok

3 agents on this tier

ATLASFORGEWARDEN

Sonnet 4.6

sonnet
Context
200k tokens
Input
$3.00 / M tok
Output
$15.00 / M tok

5 agents on this tier

PIXELVAULTDRAFTCHECKCONCIERGE

Haiku 4.5

haiku
Context
200k tokens
Input
$0.25 / M tok
Output
$1.25 / M tok

5 agents on this tier

PILOTVERSEREACHSENTINEL-1SENTINEL-2

§ How the bill adds up

Pay per token,
not per chair.

Traditional engineering teams cost in salaries — fixed monthly burn whether you ship or not. The agent stack costs in tokens — you pay for the actual work that lands.

Big-context Opus invocations are the priciest line item, used sparingly for architecture passes and security reviews. Sonnet handles the bulk of build work. Haiku runs the high-frequency, low-stakes loops — sentinels watching production, copy generation, deploy ops.

A typical platform-scale day for TripOS — schema work, a feature, a review pass, two preview deploys, sentinels watching — lands in the tens of dollars, not the hundreds. Idle days cost cents.

§ The loop · How they work together

Vision in.
Production out.

Krish posts the intent. The orchestrator breaks it into work. Builders ship. CHECK reviews. PILOT deploys. Sentinels watch. If anything looks off, the loop restarts on its own — no human ping required.

Workflow · vision → productionContinuous loop
  1. 01Krish

    VISION

    Posts intent

    “Live Mode should mirror every platform feature.”

  2. 02ATLAS

    PLAN

    Breaks it down

    Audits codebase, files tasks, picks owners.

  3. 03PIXEL · FORGE · VAULT

    BUILD

    Ships code

    Components, APIs, schema. Commits land on a branch.

  4. 04CHECK · WARDEN

    REVIEW

    Catches problems

    Type errors, regressions, security smells. Requests changes.

  5. 05PILOT

    SHIP

    Deploys

    Preview → production. Migration runs. Cache warms.

  6. 06SENTINEL-1 · SENTINEL-2

    WATCH

    Holds the line

    Errors, vitals, anomalies. Files new tasks. Loop restarts.

WATCH files new tasks back into PLAN — the loop closes itself

§ Interlinks · Who hands work to whom

The org graph,
in one read.

Every agent has a charter and a list of agents they routinely hand work to. The control plane records each hand-off; this is just the roster, flattened.

Agent

Runs on

Hands work to

  • ATLAS

    System architect

    Opus 4.7

    1M tokens

    PIXELFORGEVAULTWARDEN

  • PIXEL

    Frontend craftsman

    Sonnet 4.6

    200k tokens

    DRAFTCHECK

  • FORGE

    Backend engineer

    Opus 4.7

    1M tokens

    VAULTCHECKWARDEN

  • VAULT

    Database keeper

    Sonnet 4.6

    200k tokens

    CHECKPILOT

  • WARDEN

    Security & auth

    Opus 4.7

    1M tokens

    CHECKATLAS

  • PILOT

    DevOps

    Haiku 4.5

    200k tokens

    SENTINEL-1SENTINEL-2

  • DRAFT

    Design system

    Sonnet 4.6

    200k tokens

    PIXELVERSE

  • CHECK

    Quality control

    Sonnet 4.6

    200k tokens

    PILOTATLAS

  • VERSE

    Copywriter

    Haiku 4.5

    200k tokens

    DRAFTPIXEL

  • REACH

    SEO & growth

    Haiku 4.5

    200k tokens

    DRAFTPIXEL

  • CONCIERGE

    Runtime AI

    Sonnet 4.6

    200k tokens

    End user

  • SENTINEL-1

    Error watch

    Haiku 4.5

    200k tokens

    ATLASFORGE

  • SENTINEL-2

    Uptime & vitals

    Haiku 4.5

    200k tokens

    ATLASPILOT

§ Anatomy of a story · From your line to merged code

One story.
Six hand-offs.

A real story from this codebase — the Live Mode mirror epic — moving from a one-line vision to merged code. Every row is an event in the control plane.

Story tree · vision → mergedLive Mode mirror
  1. KRISHVISION

    EPIC · Live Mode mirror

    “Everything that is on the platform should be on Live panel mode — people are travelling and don't have laptops.”

  2. ATLASPLAN

    STORY · Audit during-trip surfaces vs. existing Live Mode

    Catalogued 35+ unwired components. Filed 4 sub-stories grouped by domain (Safety, Money, Local, Right-now).

    audit doc · 4 child stories filed

  3. PIXELBUILD

    SUB-STORY · Extract NeighborhoodsPanel + NittyGrittyPanel

    Inline panels lived in trip detail page only — moved to shared components so Live Mode can render them too.

    src/components/neighborhoods-panel.tsx · src/components/nitty-gritty-panel.tsx

  4. PIXELBUILD

    SUB-STORY · LiveCollapse helper for mobile-first disclosures

    Heavy panels (advisories, scam radar, calculators) collapsed by default so the mobile scroll stays scannable.

    src/components/live-collapse.tsx

  5. FORGEBUILD

    SUB-STORY · Wire 28 components into the new sections

    Safety / Documents / Money / Local knowledge / Right now / Plan ahead / Social / Capture / Achievements.

    src/app/app/trips/[id]/live/page.tsx · +473 / −239

  6. CHECKREVIEW

    Type-check · zero errors

    Every type was satisfied. No regressions in unrelated trip-detail flows.

  7. WARDENREVIEW

    CSP allowlist sweep

    Caught a new third-party host needing connect-src; patched the security config before merge.

  8. PILOTSHIP

    Preview deploy on every commit

    Krish opened the preview URL on his phone, eyeballed each section live.

    preview-…<branch>.app

  9. MERGED

    4 commits squashed → main · production rolled forward

    79b9f89 · 5922c24 · 77121ae · 4c57bf2

§ Control plane · The agent-native ticketing layer

There's a ticket system.
Just not for humans.

Underneath the team is a coordination layer — a control plane built for agents instead of for project managers. Every task, status update, delegation, comment, and review lives there. The agents read it; the agents write it. Krish only opens it when he wants to know what's in flight.

Agents pick up assignments, post progress as they work, hand off to the next specialist when their charter is done, and file new tickets when they spot something the human hasn't. Recurring jobs run as routines — Sentinels checking error rates every five minutes, REACH auditing meta tags weekly, VAULT verifying migrations after each deploy.

No standups. No sprints. No backlog grooming meetings. The control plane is the org chart, the kanban board, and the company chat — all consumed and produced by software.

§ Always on · 2 sentinels

While Krish sleeps,
they work.

The build team ships features. The sentinels keep them shipped. They watch error rates, deploy diffs, function panics, and Web Vitals — and open fixes before any human notices. The 3am Slack alert is a thing of the past.

SENTINEL-1

Error watch

haiku

Watches the exception stream, log streams, function panics. Opens fixes before pages.

Runs on
Haiku 4.5
Context
200k tokens
Input rate
$0.25 / M
Output rate
$1.25 / M

Pattern-match on error shapes; cheap to run continuously.

Tools

WebFetchBashTaskCreate

Hands off to

ATLASFORGE

Shipped

  • Auto-triage runbooks
  • First-fix-attempt PRs
  • Owner notification gating

SENTINEL-2

Uptime & vitals

haiku

Performance scores, vitals, deploy health, cold-start anomalies.

Runs on
Haiku 4.5
Context
200k tokens
Input rate
$0.25 / M
Output rate
$1.25 / M

Compares numbers against baselines; doesn't need a thinking model.

Tools

WebFetchBashTaskCreate

Hands off to

ATLASPILOT

Shipped

  • Vitals regression alerts
  • Deploy diff reports
  • Cold-start budget

What they watch

  • Exception stream
  • Function panics
  • Vitals regressions
  • Deploy diff anomalies

What triggers them

  • New error fingerprint
  • p95 latency spike
  • Failed cron run
  • 5xx burst above baseline

What they're allowed to do

  • Open a triage ticket
  • Draft a fix PR
  • Roll back a bad deploy
  • Notify Krish if all else fails

§ Receipts · What they shipped

The first two months,
in numbers.

Components shipped

104

React + TS, fully typed.

API routes

127

Server actions + REST.

Database models

55

Relational, indexed, normalised.

Live Mode panels

30+

Every during-trip surface.

Pitch decks

04

VC, B2B, founder, demo.

Brand assets

80

PNGs, owner-gated.

Commits to main

171

Each reviewed by the human.

Engineers on payroll

0

Yes, that says zero.

Numbers above pulled from the repo at the time of last deploy

The agents are reading this page right now,
checking it for typos.

Built by agents · supervised by Krish · 2026