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.

Read the case study →
itsm-agent — sample run (simulated)
ROUTER
UNDERSTANDING
GENERATION
REVIEW
EXECUTION
REASONING

$ ready — 14 events queued, press Run

idlesample data · deterministic replay

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.

Read the case study →
quota-governance-console — sample data
virtual clock T+0h
ProxyEnvEffective rpmStatus
checkout-apiprod-east1200TEMP · T+6h
partner-gatewayprod-west1500TEMP · T+2h
search-proxyprod-east300tier-1 · permanent
internal-billingprod-east100default · permanent
webhook-relayprod-west150tier-3 · permanent

request quota change — checkout-api

audit trail — append only

T+0hraise

partner-gateway: 6001500 rpm · j.doe

Partner batch sync window

expires T+2h

T+0hraise

checkout-api: 3001200 rpm · svc-release-bot

Flash sale — checkout surge

expires T+6h

idle — pick a proxy and submit a change

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.

Read the case study →
crowdsourcing-platform — sample data
Role
active task cap: 3 per contributorreviewer ≠ submitter (role gate)ideas auto-advance on first task pickevery action lands in the audit trail

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

✓ done

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

13:24:23 [system] COMPLETEDidea "FAQ bot for onboarding" closed — all tasks done
13:24:23 [system] SEEDEDsample platform state loaded — 1 open idea, 3 open tasks

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).

Read the case study →
observability-orchestrator — sample run (simulated)

$ ready — press Dispatch to replay the orchestration

idlesample data · deterministic dispatch

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.

Self-correcting SQL looppython
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.