Lab
A replayable console for the multi-agent pipeline behind the case studies — and the artifacts that make it honest. Sample data, deterministic execution, zero model touches.
Agent Run Simulator
Pick a question and watch it travel the pipeline: routing, parameter extraction, SQL drafting, deterministic review, execution, and narration — including the self-correcting retry loop when a draft fails schema checks.
$ ready — 14 events queued, press Run
The full RCA pipeline, including a schema failure that triggers the self-correcting retry loop before deterministic execution.
Quota Governance Console
The Apigee Accelerator's core promise, playable: raise a limit with a logged reason, advance the virtual clock, and watch the expiry sweep revert what nobody remembered to undo. Every change lands in an append-only audit trail.
| Proxy | Env | Effective rpm | Status |
|---|---|---|---|
| ▸ checkout-api | prod-east | 1200 | TEMP · T+6h |
| partner-gateway | prod-west | 1500 | TEMP · T+2h |
| search-proxy | prod-east | 300 | tier-1 · permanent |
| internal-billing | prod-east | 100 | default · permanent |
| webhook-relay | prod-west | 150 | tier-3 · permanent |
request quota change — checkout-api
audit trail — append only
partner-gateway: 600→1500 rpm · j.doe
“Partner batch sync window”
expires T+2h
checkout-api: 300→1200 rpm · svc-release-bot
“Flash sale — checkout surge”
expires T+6h
Crowdsourcing Platform Simulator
A playable miniature of the AI crowdsourcing platform. Switch between contributor and manager, pick tasks up to the cap, try to review your own work, and watch role gates, anti-gaming limits, and auto-lifecycle transitions hold the line.
Open (3)
Digest schema mapping
+25 pts · Incident digest summarizer
Tone and format rules
+15 pts · Incident digest summarizer
Rollout to #platform channel
+10 pts · Incident digest summarizer
In progress (0)
empty
In review (0)
empty
Completed (1)
FAQ corpus cleanup
+20 pts · FAQ bot for onboarding
assigned: priya.k
Role gate active: posting ideas and approving work are manager actions — switch roles to see the other side.
Leaderboard · story points
- 1.priya.kGOLD240 pts
- 2.marco.sSILVER85 pts
- 3.youBRONZE0 pts
cap note: you are at 0/3 active tasks · points are anti-gaming capped server-side
audit trail
Observability Orchestrator
Multi-agent orchestration, replayed: one LLM intent call, a deterministic router plan, and tool dispatches fanning out in parallel groups — then a verdict synthesized only from validated results. Watch who talks to the data (nobody probabilistic).
$ ready — press Dispatch to replay the orchestration
Diagnosis fan-out: four tools dispatched to isolate a latency regression and exonerate the gateway config.
Artifacts Cabinet
The mechanisms behind the claims, as code: a self-correcting SQL loop, a router prompt that asks instead of guessing, gateway quota policy, and the review checklist that gates every draft.
MAX_ATTEMPTS = 3
def generate_validated_sql(question, params, schema):
# LLM drafts. Deterministic code validates. Only then does anything run.
hints = []
for attempt in range(1, MAX_ATTEMPTS + 1):
draft = generation_agent(question, params, schema, hints)
issues = review_agent(draft, schema)
if not issues:
return draft
hints += issues
raise SQLDraftRejected(attempt=MAX_ATTEMPTS)
def execute(draft):
# Pure python. No model in the data path.
return warehouse.query(draft.sql, draft.params)The reliability lever behind the agent: drafts are validated deterministically, failures are fed back as hints, and unvalidated SQL is never executed.
How this works
- Replay, not live. The console replays a scripted transcript of a real pipeline shape with sample data — the run timing, stages, and retry behavior are simulated, the architecture is the production one.
- Deterministic execution. In the real system, SQL and API calls run in plain Python. The model drafts and narrates; it never sits in the data path. The log lines mark where that boundary is crossed.
- Self-correction. When the Review agent rejects a draft (failed schema check in the sample), the failure is returned to Generation as a hint — up to three attempts, then a clean rejection. Watch it save the run in the RCA replay.