EILAID: Engineer-In-Loop AI Development

Designed to work with the tools you already use

runtimes
- Claude Code
- Codex
- OpenCode
- Custom agents
models
- Frontier models
- Fast models
- Local models
trackers
- Jira
- Linear
- GitHub Issues
- Azure Boards
- Any MCP server
stack
- Git
- Your CI
- Slack
- Postgres + pgvector
- Docker Compose
- Helm

Why EILAID

Today, every agent starts blind.

Most teams hand agents one ticket at a time. Each one starts with no context, works with no awareness of the others, and forgets everything when the session ends.

5 agents · no shared state5 agents · 1 ledger · 1 oracle

ConflictsConflicts00

Duplicate libsLeases held05

Re-askedCited answers00

no shared record

Ledgerlive

claimPAY-412 partial refunds

leasetests/auth/refresh.spec.ts

handoffWEB-88 copy review

Oracle

Shaperno briefPAY-512

Shapersession #3PAY-512

Context: a pasted ticket
Sees: only itself
Memory: ends with session
Step: Shape → Gate 1
Lease: read-only research
Owner: @sam

ArchitectguessingBILL-37

Architectsession #4BILL-37

Context: a stale chat
Sees: only itself
Memory: ends with session
Step: Spec → split
Lease: docs/specs/bill-37.md
Owner: @maria

Builderre-reading repoPAY-412

Buildersession #5PAY-412

Context: whatever was pasted
Sees: only itself
Memory: ends with session
Step: Build
Lease: src/api/refunds.ts
Owner: @sam

Testerstarting overAUTH-221

Testersession #6AUTH-221

Context: none
Sees: only itself
Memory: ends with session
Step: Verify
Lease: tests/auth/refresh.spec.ts
Owner: @jun

Reviewerno historyWEB-88

Reviewersession #7WEB-88

Context: a diff, no why
Sees: only itself
Memory: ends with session
Step: Verify
Lease: src/web/checkout/*
Owner: @alex

Project memorydrainingcompounding62%

Poke the stage to stir things upPoke the stage to merge PAY-412

What goes wrong, every sessionThe three things every agent gets

01No context

Each session starts empty, with whatever someone pasted in.

02No awareness

Agents edit the same files and add the same library twice.

03No memory

Decisions and answers vanish when the session ends.

The loop

Two human gates. Agents do everything in between.

Every work item runs the same eight steps. The process is sized by track, so a one-line bug never pays for an epic's ceremony.

Normal feature work. A Shaper drafts the brief, and an engineer holds both gates.

Intake

PAY-412

System

PAY-412 arrives from the tracker. It is sized, given a track and a human owner, and logged in the ledger.

Produces

PAY-412track: featureledger entry

Feature track · owner @sam

Agents

A specialist for every step.

Seven personas, each with its own tools, context and output. Need another? Your team writes one as a markdown file.

Builder

step 4 · build

Implements one approved brief, test-first, inside its declared footprint. Several run in parallel, each in its own git worktree.

~/wt/pay-412$eilaid run PAY-412

◇situation report2 nearby in flight · 0 overlapping leases

›claim leasesrc/api/refunds.ts✓ granted

›write failing testtests/api/refunds.spec.ts

✗partial refund ≤ captured1 failing (expected)

›implementsrc/api/refunds.ts+48 −6

›run tests12/12✓ passing

›commita3f9c1ePAY-412: partial refunds

›open draft PR#1842PAY-412 · partial refunds

✓handoff→ Verify: Tester, Reviewer

~/wt/pay-412$

Shaper

step 2 · shape

Researches the codebase, asks the Oracle, drafts the brief.

Architect

epics · spec

Writes the spec for an epic, then splits it into features that run in waves.

Tester

step 5 · verify

Writes the failing reproduction first, then the checks that prove the fix.

Reviewer

step 5 · verify

Reviews every diff against the brief. Can run on two vendors at once.

Steward

nightly · 02:00

The night shift: dependency bumps, flaky tests, stale docs, re-indexing.

Oracle

any step · on call

Answers from the project's own knowledge, with citations. Unsure means it asks a human.

builder · PAY-512Any precedent for a swiper library?Any precedent for a swiper library?

+ gallery+ carousel+ standards3 kept ≥ 0.60

None in use yet. Standards allow a small, well-documented package, so Swiper fits. Lazy-load it.None in use yet. Standards allow a small, well-documented package, so Swiper fits. Lazy-load it.

1docs/standards.md#dependencies0.862apps/shop/package.json0.793decisions/0007-…bundle-budget.md0.74

Write your own

.eilaid/personas/*.md

A persona is a short markdown file: instructions, tools, budget and context. Commit it, review it like code, and the loop can schedule it on any runtime.

Versioned in git with everything else
Least-privilege tools, per persona
Its own budget, so it can't run away

personas/accessibility.mdnew persona

1---

2name: accessibility

3tools: [read, edit, a11y, oracle]

4budget: { minutes: 20 }

5context: [AGENTS.md, brief, tokens.md]

6output: [findings, fixes, handoff]

7---

8Check each changed view against WCAG AA.

9Fix what is inside your footprint. Raise

10a blocker for the rest. Ask the Oracle

11before adding a new dependency.

$ eilaid doctor

✓ persona valid · 4 tools · 20m budget · 3 context files

Runtimes are adapters. Pick Claude Code, Codex, OpenCode or Custom per persona in config.yaml, with a fallback.Models and runtimes

Models & runtimes

The right model for every role.

Every persona picks its own runtime and model, with fallbacks for when a vendor is down. Frontier where judgement matters, fast for chores, local when nothing may leave your network. No vendor lock-in.

Auto-demo · touch any control to take over · tap to take over

BuilderWrites the code
ReviewerReviews every diff
TesterRepros and checks
ShaperDrafts briefs
StewardNightly upkeep
OracleAnswers questions

If a vendor is down, each persona falls back to another runtime and the next tier down.

config.yaml

# .eilaid/config.yaml
preset: balanced
personas:
builder:
runtime: claude-code
model: frontier
reviewer:
runtime: [claude-code, codex]
model: balanced
tester:
runtime: claude-code
model: balanced
shaper:
runtime: custom
model: fast
steward:
runtime: claude-code
model: fast
oracle:
runtime: custom
model: fast
fallback: # when a vendor is down
runtime: codex
model: next-tier-down
network: { egress: allowed }

validsaved3 runtimes · 3 tiers · fallback on

Estimated cost per merged itemIllustrative

$8.73baselineAPI spend, per item

Builder$6.60
Reviewer$1.26
Tester$0.72
Shaper$0.08
Steward$0.05
Oracle$0.03

QualityHigh

SpeedQuick

The Oracle

Ask the project anything.

Drop in context files in whatever form you have them. Agents ask in plain English, mid-task, and every answer cites its sources. Anything the Oracle can't answer goes to an engineer, once.

eilaid ask

Answerwritten only from retrieved passages

01
Index
Every context file is split into passages, each gets a line of context, then it is embedded and stored. Re-runs on every merge.
02
Expand
A short question is widened with the asker's task, files and related terms, then written as several queries, not one.
03
Retrieve
Nearest passages from every query are merged, de-duplicated and ranked. Anything below the threshold is dropped.
04
Answer & learn
Written only from the passages, with citations. Thin evidence goes to the owning engineer once; their answer is saved.

Cost control

Every token on a budget.

Every step runs on a budget of minutes and tokens. Chores go to cheaper models, the Oracle answers repeat questions from cache, and the board shows cost per merged item. A step that blows its budget is marked stuck and sent to its owner, never retried forever.

Spend this week

illustrative

$412.8014% vs last week

Per merged item$10.86

Merged38

Stuck, caught3

Step budgets

4 running

Minutes and tokens, per step. Blow the budget and a human gets it, with no endless retries.

Parallelism you control

illustrative

Throughput8.8items/day

Spend$95/day

Median to merge11.7h

4 agents · 3 building, 1 queuedOverlapping work is queued by leases

Smart savings

−68%per run

Steward on a fast model

Dependency bumps, flaky tests, stale docs. 22 runs, $9.00 this week.

71%hit rate

Oracle answers from cache

912 repeat questions answered free this week, keyed by question and commit.

Saved this week −10%≈ $44.57

Guardrails

Daily cap $120New steps wait for tomorrow
Per-item cap $15Item pauses for its owner
Alert at 80% $96Ping on phone or Slack
Pause on anomaly 3× spikeSpend spike stops new runs

Today$35.30 / $120

All guardrails on

Remote monitoring

Watch every agent. From anywhere.

A live board of every agent and step, with heartbeats, logs and budget used. When something needs a person, it finds you, and one tap puts the agent back to work.

Runs

Live

All runs 5Needs you 0

notifyPhoneSlackTracker

ItemStatusHeartbeatBeatTimeTokens

Session refreshAUTH-221 · 5/8 VerifyVerifying

18:40742k

Rotate signing keysAUTH-226 · 2/8 ShapeShaping

6:1288k

Checkout copyWEB-88 · 5/8 VerifyReviewing

6:4892k

Partial refundsPAY-412 · 4/8 BuildBuilding

31:051.18M

Image galleryPAY-512 · 4/8 BuildBuilding

18:22516k

14:02:04builderlease ok · src/web/gallery/** (footprint)

14:02:09builderwrite tests/web/gallery/lazy-load.spec.ts

14:02:15buildervitest gallery → 3 failing (expected)

14:02:18oracleQ: any precedent for a swiper library?

14:02:19oracleA: cached · Swiper ok, lazy-load it [3]

14:02:26builderpnpm add swiper@11 · lazy-loaded

ledger5 agents live · heartbeat every 5s

EILAIDAUTH-226now

Gate 1 · brief ready for review

Rotate signing keys, by Shaper

EILAIDWEB-88now

Question from Reviewer

Is there a copy style guide?

EILAIDAUTH-221now

Tester hit its token budget

800k of 800k used → needs you

Multi-engineer teams

Many engineers. Many agents. Zero collisions.

Ownership routes every gate and question to the engineer who owns the code. Leases stop agents and people colliding, and everyone works from one shared ledger.

Codebase mapmain

payments48@samSM

refunds.ts

summary.ts

ledger.ts

copy.ts

webhooks.ts

currency.ts

+42+43

billing31@mariaMK

invoices.ts

queue.ts

worker.ts

plans.ts

+27+29

auth22@junJL

session.ts

tokens.ts

refresh.ts

session.test.ts

oauth.ts

+17+20

web/gallery27@alexAR

Gallery.tsx

Lightbox.tsx

lazy.ts

thumbs.ts

+23+25

search19@alexAR

index.ts

query.ts

ranking.ts

synonyms.ts

+15+17

db/migrations64@samSM

0047_refunds.sql

0046_invoices.sql

0045_tax.sql

+61+62

api36@junJL

routes.ts

refunds.ts

errors.ts

+33+34

notifications14@mariaMK

email.ts

templates/

push.ts

+11+11

PAY-412builder

WEB-88builder

AUTH-221tester

BILL-37builder

PAY-512shaper

ledgerEvery claim, lease, question and gate#1,284

14:02queueWEB-88 queued behind PAY-412 · overlaps payments/checkout/summary.ts@alex
14:01askPAY-512 Extend Lightbox or replace it? → owner of web/gallery@alex
14:00leasePAY-412 leased 5 files across payments api db@sam

Leases, not locksoverlaps wait their turn
Routed to the ownergates and questions
One shared ledgersame view for everyone

Continuous development

It never stops shipping. It never stops learning.

The same loop runs features, bug fixes, migrations, nightly maintenance and learning. Only the trigger and the track change, and every rejection makes next week’s work smoother.

Always onOne day on the board, replayed

09:24

It gets better every week

Illustrative

Escalations per 10 items1.1~~3.6~~−69%

Review rounds per change1.4~~2.8~~−50%

Rules, tests and context learned, cumulativehover to read a week

Learned today

0/5 approved

Every rejection, review comment and repeated question becomes a file in git, approved by its owner.

+testSession refresh regression testtests/auth/session-refresh.spec.tsfrom AUTH-221 incident · approved by @jun11:15
+ruleMoney values use the Money type.eilaid/rules/money-type.mdfrom Review round on PAY-412 · approved by @sam12:51
+contextGallery image sizes, answered once.eilaid/context/gallery-images.mdfrom Routed question on PAY-512 · approved by @sam14:25
+ruleLazy-load carousel libraries.eilaid/rules/lazy-load-carousels.mdfrom Gate 2 note on PAY-512 · approved by @sam17:13
+testPDF export timeout reproductiontests/billing/pdf-export.spec.tsfrom Error report · approved by @maria22:18

Platform

Your tracker. Your agents. Your servers.

A built-in tracker designed for agents, adapters for the one you already use, any agent runtime, and it runs on a laptop, inside your network or hosted.

Trackersboth ways

Agent runtimesone adapter

Infra & notify

EILAID coreledger.* · oracle.* over MCP

Deployment

Runs on a laptop, inside your network, or hosted.

All three run the same code on the same data model. Start on a laptop and move to a server without migrating anything.

Local

A single binary: CLI, terminal board and a local web UI.

~/payments-api

$ eilaid init

✓created.eilaid/config.yaml, personas/

✓ledgersqlite · .eilaid/ledger.db

✓oracleindexed 1,284 passages

→nexteilaid board

Storage: SQLite and files in the repo
Best for: One developer or a small team trying it out

Self-hosted

Docker Compose or a Helm chart, inside your network.

infra/eilaid

$ docker compose up -d

✓corerunning · sso enabled

✓postgrespgvector ready

✓runners4 sandboxed

✓modelsprivate endpoint

$ helm install eilaid ./eilaid

✓releaseeilaid deployed

✓podscore, oracle, runners ×4

✓postgrespgvector ready

✓egressinside your network

Storage: Postgres with pgvector
Best for: Data residency, single sign-on, audit needs and local models

Cloud

A hosted service. Sign in and connect your repo and tracker.

~/payments-api

$ eilaid login

✓signed in@sam

✓workspacepayments · own tenant

✓linkedrepo + tracker

→nexteilaid board

Storage: Managed, per tenant
Best for: Teams that want nothing to operate

Same code, same data. Move between modes without a migration.

Agents get the least access they need.

Least-privilege toolsEach persona gets its own tool allowlist.
Sandboxed runnersOne persona per step, in an isolated worktree.
No production secretsRunners never hold prod credentials.
Every action in the ledgerClaims, edits and decisions, append-only.
Only humans mergeAgents propose. An engineer merges to main.

Principles

Nine rules behind the loop.

Tracks, personas, gates and the Oracle all fall out of these. When in doubt, an agent (or an engineer) comes back here.

methodology.mdRead them in full

01
Engineers own intent and acceptance; agents own execution.

Humans sit at gates, not inside every loop.
Diagram: a ticket moves through the eight loop steps. Agents run Intake, Shape, Build, Verify, Merge and Learn. An engineer appears only at Gate 1, to approve the brief, and at Gate 2, to accept the change. A failing check loops back to Build with no human involved.
02ledger
- 09:41claimPAY-412builder-2
- 09:42questionAUTH-221oracle
- 09:43decisionBILL-37@maria
- 09:44handoffPAY-412tester
State lives outside agents.

If it is not in the ledger or the repo, it did not happen.
03
Context is assembled, not accumulated.

Each step starts with a fresh context window, built for that step.
Brief, footprint, decisions and cited answers. Nothing stale.
04
Verification is the contract.

Every acceptance criterion maps to an automated check.
A criterion nobody can check is an open question, not a requirement.
05main a3f9c2e
- ├─briefs/PAY-412.mdM
- ├─personas/builder.mdM
- ├─decisions/0007-swiper.mdM
- └─knowledge/payments.mdM
Everything is a file in git.

Briefs, personas, decisions and knowledge are plain files.
06
Ask, don't guess.

Unsure? Ask the Oracle, then a human. Never invent an answer.
Each human answer is saved, so it is only asked once.
0712 weeks
Every failure compounds.

Each mistake becomes a rule, a test, a lint check or a skill.
08
Right-sized process.

A typo fix must not need a requirements document.
Patch, Feature or Epic: the track fits the work.
09config.yaml
Vendor-neutral.

Trackers, agent runtimes and models are swappable adapters.

Close the loop.

Start with the method: a brief, two gates and a shared ledger. Add the tools when you need them. Everything is plain files in git, so nothing is locked in.

Get started Read the methodology

npx eilaid init

proposed CLI

Engineers decide.
Agents build.
The project remembers.

Engineers decide. Agents build. The project remembers.

01No context

02No awareness

03No memory

Intake

Index

Expand

Retrieve

Answer & learn

Runs on a laptop, inside your network, or hosted.

Engineers own intent and acceptance; agents own execution.

State lives outside agents.

Context is assembled, not accumulated.

Verification is the contract.

Everything is a file in git.

Ask, don't guess.

Every failure compounds.

Right-sized process.

Vendor-neutral.

Close the loop.