8-Agent Pipeline · Real-Time Streaming

Software Architecture
Review, Automated

Submit a binary architecture decision. Eight independent agents reason in parallel, surface conflicts, and produce an auditable recommendation, streamed live.

Run Live Review View Architecture

Parallel Execution

4 specialist agents run concurrently via ThreadPoolExecutor, cutting wall-time vs sequential by ~60%.

Explicit Conflicts

The Detector surfaces where specialists disagree: on recommendations, confidence, or risk reasoning.

Policy-Gated Approval

The Gate blocks approval if the Critic flags issues or the Detector finds irreconcilable conflicts.

Submit Architecture Decision

Agents reason independently; they cannot see each other's outputs

Agent Pipeline

Planner
gpt-5-mini
↓
parallel
Scalability
gpt-5
Security
gpt-5
Cost
gpt-5
Maintain.
gpt-5
↓
Detector
gpt-5-mini
↓
Critic
gpt-5-mini
↓
Synthesizer
gpt-5.5
↓
Gate
rule-based
Submit a decision to observe agent reasoning in real time

Architecture

Deterministic state machine with typed agent I/O and policy-gated approval

LangGraph Orchestration

Two compiled graphs: parallel (fan-out/fan-in on the 4 specialists) and sequential (baseline for benchmarking). Shared flat TypedDict state eliminates merge conflicts in parallel execution.

Pydantic I/O Schemas

Every agent input and output is a typed Pydantic model. Planner writes named sub-questions per specialist; no generic list indexing. OpenAI structured outputs validate JSON schema at the model layer.

WebSocket Streaming

LangGraph astream_events(v2) fires on_chain_start and on_chain_end per node. Specialists arrive out of order as threads complete; the UI renders each independently.

The 8 Agents

Each agent sees only what it needs; specialists cannot see each other's outputs

PLANNER · gpt-5-mini

Planner

Decomposes the question into one targeted sub-question per specialist. Extracts key constraints that all agents must respect.

SPECIALIST · gpt-5

Scalability

Throughput, horizontal scaling limits, sharding, connection pool exhaustion, data growth degradation.

SPECIALIST · gpt-5

Security

Attack surface, auth boundaries, encryption at rest/transit, CVE exposure, compliance posture, managed vs self-patched.

SPECIALIST · gpt-5

Cost

TCO, licensing, egress fees, operational overhead, scaling cost curves, step-function pricing jumps.

SPECIALIST · gpt-5

Maintainability

Team familiarity, debugging story, migration risk, vendor lock-in, operational toil, reversibility.

DETECTOR · gpt-5-mini

Detector

Surfaces recommendation splits, confidence conflicts, and risk contradictions. Classifies each as blocking or non-blocking.

CRITIC · gpt-5-mini

Critic

Challenges specialist assumptions. Surfaces collective blind spots. Sets requires_revision if issues would change the outcome.

SYNTHESIZER · gpt-5.5

Synthesizer

Weighs all evidence and commits to one option. Resolves conflicts. Lists tradeoffs and unresolved risks. Never outputs "neutral".