AgentGrid

Greenfield horizontally scaled AI platform with two verticals on shared infra:

1. Autonomous coding agents — plan → isolated workspace → patch → pytest verifier → merge status (multi-agent retry vs single-agent baseline).
2. Behaviour analytics — synthetic trading/research product journeys (funnels, anomalies, evidence-grounded NL insights) with consent + deletion.

Placement pitch: I can ship distributed agent systems with evals and a real analytics product surface—not a chat demo.

No third-party API keys required. Planners and workers use offline stubs; CI and local demos never call OpenAI/Anthropic/etc. Local auth uses the documented demo token dev-token (see .env.example).

Repo: https://github.com/Wojtek-06/AgentGrid · License: MIT
Non-goals: Fitness-App code/domain; live extractive agents against private remotes; LLM network calls in CI.

---

Architecture

Full write-up: docs/ARCHITECTURE.md.

                    ┌─────────────────────────────────────┐
                    │         Dashboard (static UI)         │
                    │   board · metrics · funnel · SSE      │
                    └──────────────────┬──────────────────┘
                                       │ Bearer token
                                       ▼
┌──────────────┐   enqueue    ┌────────────────┐   dequeue   ┌─────────────────┐
│ FastAPI API  │─────────────►│ Queue (Redis   │────────────►│ Coding workers  │
│ jobs · eval  │              │  or in-proc)   │             │ plan→patch→test │
│ analytics    │◄─────────────│                │◄────────────│ + merge path    │
└──────┬───────┘   status     └────────────────┘   results   └────────┬────────┘
       │                                                              │
       │                              ┌───────────────────────────────┘
       ▼                              ▼
┌──────────────┐              ┌─────────────────┐
│ SQLite / PG  │              │ .artifacts/     │
│ jobs·events  │              │ patch + checklist│
└──────────────┘              └─────────────────┘

Dogfood issues (local sandboxes, QuantForge/ChainVenue-shaped):

Issue ID	Story	Eval contrast
qf-leakage-guard	Look-ahead mid bug	multi retries, single fails
qf-ewma-alpha	EWMA weights swapped	multi retries, single fails
cv-basis-bps	Wrong basis sign	both modes can fix

Published eval numbers

Source of truth: data/eval_results.json (regenerate with python scripts/run_eval.py).

Mode	Success rate	Avg tokens
Single-agent	33% (1/3)	~200
Multi-agent	100% (3/3)	~287

Per-issue breakdown and interview script: docs/EVIDENCE_PACK.md.

---

Quick start

A separate worker process needs a shared Redis queue (in-process queue is for pytest only).

Redis (pick one):

• WSL: sudo service redis-server start (or redis-server) on localhost:6379
• Docker: docker compose up -d redis

cd C:\Projekty\Quant\AgentGrid
python -m pip install -r requirements.txt

# Terminal A — API (sets USE_REDIS=1 + demo token)
.\scripts\run_api.ps1

# Terminal B — coding worker (heartbeats on /api/health)
.\scripts\run_worker.ps1

Tests (no Redis required): $env:PYTHONPATH="backend"; python -m pytest -q

60-second demo

1. Open http://127.0.0.1:8000 — token dev-token; health strip should show redis on · workers ≥ 1.
2. Select issue qf-leakage-guard, mode single → Enqueue → status goes failed.
3. Same issue, mode multi → Enqueue → status goes succeeded (retry path).
4. Click Load published (instant) or Run eval → multi 100% vs single 33%.

Optional: python scripts\seed_analytics.py then Refresh for the research funnel.

---

Auth

Protected routes expect:

Authorization: Bearer <AGENTGRID_API_TOKEN>

Default token is dev-token (see .env.example) — change it outside local demos.
SSE (EventSource) cannot set headers, so only /api/jobs/stream accepts ?token= (Bearer everywhere else).
Health (GET /api/health) is open and reports queue depth, Redis reachability, and live worker heartbeats.

---

Docker (optional)

# API + Redis + one worker
docker compose up --build

# Horizontal scale story — extra workers share the Redis queue
docker compose up --build --scale worker=2

Optional Postgres profile (SQLite remains CI/default):

docker compose --profile postgres up --build api-pg worker-pg postgres redis

Extras:

python scripts\run_eval.py          # refresh data/eval_results.json
python scripts\seed_analytics.py    # demo funnel + retention

---

API

All rows except health require Authorization: Bearer <token> (or ?token= on SSE).

Method	Path	Auth	Purpose
GET	/api/health	—	Liveness, queue, Redis, worker heartbeats
GET	/api/jobs/issues	Bearer	Dogfood catalog
POST	/api/jobs	Bearer	{issue_id, mode, idempotency_key?}
GET	/api/jobs	Bearer	Board snapshot
GET	/api/jobs/stream	Bearer or ?token=	SSE live board
GET	/api/jobs/{id}	Bearer	Job detail + patch/log
POST	/api/jobs/{id}/cancel	Bearer	Cancel queued/running
POST	/api/jobs/{id}/retry	Bearer	Re-enqueue failed/cancelled
GET	/api/eval/latest	Bearer	Published data/eval_results.json
POST	/api/eval/run	Bearer	Re-run multi vs single harness
GET	/api/metrics/overview	Bearer	Tokens / $ / latency / queue
POST	/api/analytics/events	Bearer	Batch ingest
GET	/api/analytics/funnel	Bearer	Funnel + anomalies + insight
GET	/api/analytics/retention	Bearer	Day-1 retention cohorts
GET	/api/analytics/operator-funnel	Bearer	Operator telemetry funnel
POST	/api/analytics/consent	Bearer	Consent flag
DELETE	/api/analytics/users/{id}	Bearer	Erase + block

Responses include X-Request-ID (echo client header or generate). API + worker logs use structured request_id=… fields.

---

Status vs placement plan

Deliverable	Status
Coordinator + workers + queue + verifier	Done (local/Redis)
Multi vs single eval	Done (scripts/run_eval.py → data/eval_results.json)
Merge artifacts + human review checklist	Done
Merge conflict risk surfacing + tests	Done
Cancel / retry / queue backpressure	Done
Observability metrics + request IDs / structured logs	Done
SSE live job board	Done
Analytics + privacy	Done (research journeys)
Operator telemetry + retention cohorts	Done
Optional Postgres compose profile	Done (SQLite remains CI default)
Horizontal scale story	Documented + compose --scale worker=N
Dogfood on QF/CV-shaped issues	3 local sandboxes
Evidence pack / demo video	Docs ready; video user-owned

Sibling status: docs/PORTFOLIO_STATUS.md
Docs: docs/EVIDENCE_PACK.md · docs/THREAT_PRIVACY.md · Screenshots: docs/images/

---

Sibling projects

• QuantForge — C++ LOB MM lab (dogfood-shaped issues)
• ChainVenue — Foundry CLOB–AMM lab (dogfood-shaped issue)