The platform.

One cockpit, parallel sessions, and a record that compounds.

THE PLATFORM

Coding agents that don't melt your context window.

Not a plug-in — a platform. Everything an agent touches becomes an artifact: inspectable, citable, reusable.

RUN LIVEcore-billing · lane kai/0148-stripeelapsed 04:12
9%CONTEXT
47ARTIFACTS
$12.40RUN COST
BACKEND

writing stripe-usage.ts

58%
TESTS

41/58 integration passing

RUNNING
BROWSER

verifying checkout flow in live DOM

34%
✓ commit 6e315c1 · PR #612 opened · verified against HEAD — No merge. No deploy. No scope widening.
Artifacts, not tokens

Big files and tool outputs live on disk. The model cites them on demand — context stays lean, runs finish.

Memory that compounds

/Projects plus a knowledge graph. Every run starts with everything the last one learned.

Parallel by default

Specialised agents spawn mid-run with context distributed between them — not duplicated.

PARALLEL SESSIONS

Run five things at once. Lose track of none.

Every session gets its own branch and its own record. They don’t share context — they share memory.

Run 1WO-0148 · ⎇ backendKiwi Code ▾✕
Ship the metered billing slice. Don’t touch subscription paths.
What I did
1 file · commit 6e315c1
PR #612 opened · claims match HEAD ✓
No merge. No deploy. No scope widening.
backend is working ▍
› Message backend…
Run 2WO-0149 · planningClaude ▾✕
Tests — pick one
A · focused idempotency file (preferred)
B · extend closed-loop suite
Blocked on you.
A
Lane cut · ⎇ kai/0149-oauth-v2
plan locked · 5 steps ✓
planner is working ▍
› Message planner…
Run 3WO-0151 · ⎇ testsKiwi Code ▾✕
Kill the flaky retry test for good.
Pinned seed 42 · root cause: shared state
backoff fix pushed · suite green × 20 runs
re-verified ✓
tests are running ▍
› Message tests…
+Add runa new agent, full context, instantly
THE RECORD

One shared memory. Every decision on file.

SHARED MEMORY
Stripe webhook pattern
written by Security · approved by Marcus L
cited · Run 1
Never store raw payment payloads
written by Priya S (human)
cited · Run 1cited · Run 3
Sync invoice generation — failed 2024
postmortem · kills bad proposals on sight
Write once, never re-explain. Every run starts with everything the last one learned.
Humans and agents share one canon. Your ADR guides an agent’s call six months later.
Failures compound too. A postmortem blocks the same bad idea in every future run.
DECISION TRACE
“New metering service, or extend the event processor?”
A1New standalone service — clean, but duplicates idempotency + audit
A2Sync handler in webhook path — reproduces the 2024 timeout failure
A3Extend the processor — inherits retries, audit log, canary for free
deciding…
Every call has a why. Alternatives considered, on the record — not just the output.
Debuggable judgment. When something ships wrong, read the reasoning, not the diff alone.
Trust is earned with data. Stated confidence vs. how it actually held up, per agent.
THE CODE FACTORY · TEAMS

Where runs become work orders.

On Teams, sessions roll up into a factory: a board your org can read, gates you configure, and a record you can audit.

kiwicode.ai · factory / work orders
CODING · 2
WO-0147P2
Payout approvals dashboard
⎇ kai/0147-payouts
WO-0151P2
Webhook retry queue
⎇ kai/0151-retry
TESTING · 1
WO-0148P0
Stripe usage-based billing
⎇ kai/0148-stripe
REVIEW · 1
WO-0144P0
KMS key rotation
⎇ kai/0144-kms
EXITED · 7 TODAY
WO-0141EXIT 0
Tax-rate sync job
merged · 11:04
WO-0138EXIT 1
Ticket auto-categorize
killed · lesson → memory
Failures stay on the board — killed with a full trace, and the lesson writes itself back to shared memory.
factory / lanes — git truth
WOLANEHEADPRRECON
0148kai/0148-stripe6e315c1#612MATCH
0147kai/0147-payouts4211e0a—MISMATCH
0144kai/0144-kms878228b—MATCH
0143main · canary789a297#609MATCH
read from worktrees — never from agent self-reports
factory / gates — your call, per surface
Payment surfaces → human security sign-offON
Prod credentials → owner approvalON
Merge to main → human reviewON
Architecture → agents decide · calibration 91%OFF
Staging deploys → canary + auto-rollbackOFF
Autonomy is a dial, not a mode — remove gates as the calibration record earns it.
THE RUN LIFECYCLE

From intent to verified handoff.

Work order

Requirement, constraints, acceptance criteria — the contract the run is judged against.

Plan

The agent reads /Projects memory first — zero tokens re-explaining your codebase.

Lanes

Parallel sessions, each on its own branch. They share memory, never context.

Reconcile

Claims diffed against git truth — HEAD, dirty state, PR status — deterministically.

Handoff

Branch · SHA · PR · what was explicitly not done. Machine-checkable, every run.

MEMORY & CONNECTORS

What lives in .kiwi/

Every workspace carries its own metadata — the compounding state that makes run N+1 smarter than run N.

CODEBASE_SIMPLIFIED.md

The living map of your repo — architecture and entry points, kept current so no run starts blind.

decisions.md

Every choice and its why, recorded at decision time. The trace your review reads.

Tasks.py

Open work, structured — what the next run picks up without being told.

Completed.py

The shipped record — what landed, where, and when. Memory of done.

100+ connectors, all server-side OAuth — credentials never enter model context.

NetSuiteOpenRouterGitHubSlackGoogle DriveGmailStripePostgresNotionJiraLinear+ 90 more

Stop feeding the context window.
Start shipping verified work.

Kiwi Code is onboarding enterprise teams now.

› Contact Us
or ask us anything — abhishek@meetkiwi.co