Asadullah Shafique

A multi-agent system has to answer four questions: who runs next, what happens when a specialist finds something, what happens when one fails, and how you prove afterwards what actually occurred. The common answer puts all four inside a prompt, which makes the most important behaviour in the system untestable without a live model, non-deterministic across runs, and unprovable to a reviewer.

All four move into typed state and plain Python on a LangGraph supervisor. Specialist agents (RiskModeler, ComplianceChecker, ClientAdvisor, Reporter) return only to the supervisor and never to each other, so every routing decision sits in one function that can be tested in isolation. Six invariants are enforced structurally rather than by convention: only the supervisor routes, required specialists always run before a report, every run terminates, every hop is recorded on success and failure, routing never consults a model, and concurrent runs are isolated by run id so breaker state cannot leak between them.

41 tests pass with no API key and no network, because routing is model-free — that is what makes 'the supervisor never skips ComplianceChecker' a millisecond CI assertion rather than a claim in a README. A looping supervisor needs a hop budget, not just a retry cap: retries bound one node's failures, but only a budget bounds the cycle. Escalation is graduated and isolated in one function marked POLICY SEAM — halt on critical, one bounded mitigation round on high — so an organisation's risk posture is a single-function edit. Ships a CLI, a FastAPI service whose GET /trace/{run_id} lets a run be interrogated after the fact, a Dockerfile, and CI across Python 3.11 to 3.13.

Anthropic's MCP and Google's A2A solve adjacent problems, and a real enterprise agent needs both: MCP to call tools you host over stdio, A2A to delegate work to agents another company operates over HTTP. Wiring them together ad hoc produces N-squared translators and, worse, drops the governance context at every seam. The caller's identity, the data's sensitivity label and the correlation id all evaporate the moment a message changes protocol.

One normalised ProtocolEnvelope that every message is lifted into and lowered out of, so adding a protocol costs a single codec pair instead of N-squared adapters. Governance rides inside the message rather than in transport headers, because headers do not survive an MCP stdio hop. Three agents sit on a LangGraph state machine: MCPConnector binds external tools, A2AGateway delegates across vendor boundaries, and AuditAgent inspects twice per hop, once before dispatch and once after, because ingress cannot know a response will carry a national ID and egress cannot un-send a request. Both protocols are implemented to the wire rather than simulated: real JSON-RPC over stdio in a subprocess, and a real HTTP peer serving an Agent Card at /.well-known/agent.json.

41 tests pass with no API key and no network. Every hop lands in a hash-chained ledger that covers a payload digest rather than the payload, so the integrity proof survives dropping the sensitive bytes. The live demo replays six real traces and recomputes all ten SHA-256 digests in the visitor's own browser, so tamper detection can be checked rather than taken on trust. Enforcement is isolated in one function marked POLICY SEAM, which makes an organisation's risk posture a single-function edit.

Pakistani developers lack a local, context-aware Q&A platform. Most alternatives are too generic and not community-driven.

Open-source community platform with threaded Q&A, blogs, project collaboration, and AI-powered answer suggestions. Built with Next.js 15 App Router and shadcn/ui.

Live, with source and a running deployment. This site is that codebase: the Agents SDK orchestrator, the MCP server and the constitution all ship from this repository.