Live Demo · Agentic Loop Engineering AAROP An autonomous multi-agent reasoning system. Watch objectives flow through the full agentic loop — Perceive → Plan → Act → Observe → Reflect → Adapt — with delegated worker agents, self-verification, resilient recovery, bounded autonomy, and a replayable trace.
Built by Devayan Mandal · github.com/devtechedge/aarop
1 · Live Agentic Loop Run Agent ▶
🧮 Compute 📚 Research + synthesize (multi-agent) 🏗️ Design task (multi-agent) 💥 Failure → retry → escalate ⛔ Budget exhaustion → escalate
Perceive
Normalize input + recall memory (RAG)
→
Plan
Build a cost-aware task graph
→
Act
Invoke validated, sandboxed tools
→
Observe
Capture results + detect anomalies
→
Reflect
Critic verifies against acceptance criteria
→
Adapt
Replan / retry / escalate to human
Execution Trace (replayable) Run a scenario above to see the live trace…
2 · Multi-Agent Orchestration The Orchestrator decomposes the objective and delegates to specialized worker agents. Agents light up as they engage during a run.
Orchestrator
Owns the global plan & delegates
Researcher
Web / KB / RAG retrieval
Coder
Code & computation execution
Analyst
Data synthesis & analysis
Verifier
Critic — gates every result
Memory
Episodic / semantic recall & commit
3 · System Architecture Production reference stack — every demo component maps to a real backend.
API Gateway · auth · budget · schema validation Orchestrator (Agentic Loop) Workflow Engine (Temporal) Worker Agents vLLM / OpenAI Tool Registry sandboxed Memory pgvector / Qdrant Model Router cloud + self-host Observability (OpenTelemetry) · Eval Harness · FinOps
Full C4 diagrams, sequence diagrams & 5 ADRs in ARCHITECTURE.md .
4 · Engineering Rigor Bounded Autonomy Step / cost / time budgets; escalates to a human instead of looping forever. Try the budget scenario above.
Resilient Recovery Tool retries, circuit breaker, anomaly detection in Observe. Try the failure scenario above.
Self-Verification A critic agent scores every result against acceptance criteria before commit.
Multi-Agent Orchestrator delegates to Researcher / Coder / Analyst / Verifier / Memory agents.
Observability Structured, replayable trace per run — exportable as JSON.
Tested Core 9 tests, 92% coverage on the Python core; CI runs Python 3.10–3.12 + the Next.js build.