# Agent architecture and orchestration > I design multi-agent systems that do real work, not demos. Supervisors, orchestrator-worker patterns, tool use, MCP and human-in-the-loop review, built on LangGraph, the Claude Agent SDK, Google ADK and Copilot agents. I decide what an agent should do, what deterministic code should do instead, and where a person signs off. - Core skills: Multi-agent orchestration, Claude Agent SDK, LangGraph, Model Context Protocol, Realtime voice agents, Copilot custom agents and skills - Domains: Telecom, Banking and financial services, Quality engineering, Legacy modernization, Conversational and voice AI, Developer platforms and tooling - URL: https://farhanbaluch.com/capabilities/agent-architecture ## Case studies - [Multi-agent legacy documentation](https://farhanbaluch.com/work/legacy-docs-agents.md) (Aug 2026 – Oct 2026, Client demo, embedded systems): 16 sections, every claim verified against cited lines. Modernizing with AI starts with understanding the old system, and an LLM summary that hallucinates is worse than none. Deterministic analysis, explorer agents and an independent reviewer together produce documentation people can trust, and that becomes context for the agents that do the migration. - [Agentic legacy modernization and code intelligence](https://farhanbaluch.com/work/legacy-modernization.md) (2025 – 2026, Enterprise bank): 70%, shorter modernization timeline. Spec-driven agents transform the code and a multimodal agent verifies migrated reports, while a GraphRAG platform gives engineers and agents a queryable map of the estate. - [Estimation platform for an enterprise bank](https://farhanbaluch.com/work/estimation-platform.md) (2025, Enterprise bank): 80%, less manual estimation effort. Agents research and draft the estimate, a regression model anchors the number, retrieval over past projects supplies citations, and an expert approves. The AI does the slow part and a person keeps the decision. - [Realtime voice agent for banking onboarding](https://farhanbaluch.com/work/banking-voice-onboarding.md) (Dec 2024, Enterprise bank, proof of concept): Designed and built the agent. Voice is where agents meet customers with no screen to fall back on. A realtime speech-to-speech agent with a confidence-based handoff keeps the customer moving without letting the model guess on a regulated process. - [Autonomous testing for a conversational voice agent](https://farhanbaluch.com/work/voice-agent-testing.md) (Jun 2026 – Jul 2026, National telecom carrier): <10 min, to generate branch-covering test cases for a call flow. Agents testing an agent. Worker agents hold real multi-turn conversations with a production voice assistant, a design agent writes the test cases from call-flow designs, and an LLM judge scores meaning rather than wording. - [Custom Copilot agents for test-automation teams](https://farhanbaluch.com/work/copilot-agent-program.md) (Mar 2026 – Sep 2026, Enterprise test-automation teams): 27, agents and skills catalogued across seven SDLC phases. Teams had AI assistants, but generic prompts wrote generic tests that ignored each team's frameworks, page objects and naming, so adoption stayed shallow. Encoding those conventions into agents and skills is what made AI-generated tests mergeable, and the catalog is what turned individual agents into a toolkit the whole programme reuses. - [Agent and skill distribution platform](https://farhanbaluch.com/work/agent-distribution.md) (May 2026 – present, National telecom carrier): 18 repos, on versioned agent packages, from 7 packages. Agents only change how an organization works if every team runs the same current, reviewed version. This is the supply chain for the AI programme: agents and skills versioned like dependencies, generated from one registry, and installable on locked-down laptops. ## Evidence from the log - 2026-10-07: **Demoed multi-agent documentation for legacy firmware.** An orchestrator splits the codebase across explorer agents, and an independent reviewer agent rejects any claim it can't trace to a cited line. AI angle: Shows an agent system whose output an engineer can trust, because a second agent verifies every claim against the source before it reaches the report. [Agent architecture and orchestration; AI evaluation and quality] (https://farhanbaluch.com/work/legacy-docs-agents.md) - 2026-09-22: **Launched a searchable catalog of 27 agents.** Mapped every agent and skill to one of seven SDLC phases and published them in a docs portal, so engineers find an agent before they write a new one. AI angle: Turns a pile of agents into a toolkit. Reuse rises, duplicates fall, and the programme can see which SDLC phases still have no agent support. [AI adoption at enterprise scale; Agent architecture and orchestration] (https://farhanbaluch.com/work/copilot-agent-program.md) - 2026-09-15: **Added a reviewer agent to the documentation pipeline.** Early drafts read well but overstated what the code did. A separate reviewer agent now checks each section against the source and sends unsupported claims back. AI angle: Generator-reviewer separation applied to hallucination. The same model is never trusted to check its own work. [AI evaluation and quality; Agent architecture and orchestration] (https://farhanbaluch.com/work/legacy-docs-agents.md) - 2026-08-25: **Moved agent distribution to a single registry.** Copies of agent files had drifted across repositories. Now one registry generates every manifest, and packages are versioned and reviewed before release. AI angle: Agents are now shipped like software. A prompt or skill fix reaches every team, and nothing changes agent behavior without review. [AI adoption at enterprise scale; Agent architecture and orchestration] (https://farhanbaluch.com/work/agent-distribution.md) - 2026-08-18: **Started the legacy documentation agents.** Modernization projects stall because nobody understands the old code. Began an agent pipeline that combines deterministic static analysis with parallel LLM explorers. AI angle: Deterministic tools handle what must be exact (call graphs, registers) and agents handle what needs reading and explanation. The result becomes context for the agents that do the migration. [Agent architecture and orchestration; Knowledge, retrieval and context] (https://farhanbaluch.com/work/legacy-docs-agents.md) - 2026-08-13: **Changed the rollout plan for code-review skills.** Found that code-review skills only load from a repository's default branch, so teams testing on feature branches saw nothing. Rewrote the rollout steps around that. AI angle: Knowing how the agent platform actually loads skills is the difference between a rollout that works and teams deciding AI review "doesn't do anything". [AI adoption at enterprise scale; Agent architecture and orchestration] (https://farhanbaluch.com/work/copilot-agent-program.md) - 2026-07-07: **Generated test cases from call-flow designs in under 10 minutes.** A Claude tool-use agent reads a call-flow design and writes one test case per branch, checked against a 22-case gold standard. AI angle: An agent does the slow design work a tester used to do by hand, and a gold standard keeps it honest. [Agent architecture and orchestration; AI evaluation and quality] (https://farhanbaluch.com/work/voice-agent-testing.md) - 2026-06-15: **Architected autonomous testing for a production voice assistant.** Manual testing meant calling the assistant again for every branch. Designed an orchestrator that dispatches parallel worker agents to hold multi-turn test conversations. AI angle: Agents testing an agent. Worker agents play the customer, so coverage scales with compute instead of tester hours. [Agent architecture and orchestration; AI evaluation and quality] (https://farhanbaluch.com/work/voice-agent-testing.md) - 2026-04-20: **Rolled out custom Copilot agents to test-automation teams.** Wrote agents and skills that generate Playwright tests in each team's own conventions, and ran hands-on sessions so teams actually used them. AI angle: Generic assistants write generic tests. Encoding each team's frameworks and conventions into agents is what made the output mergeable, and the sessions are what made people use it. [AI adoption at enterprise scale; Agent architecture and orchestration] (https://farhanbaluch.com/work/copilot-agent-program.md) - 2026-03-23: **Generated contract and API tests with agents.** Extended agent workflows beyond UI tests to Pact contract tests and API suites, so service changes are caught before end-to-end runs. AI angle: Moves AI test generation down the test pyramid, where tests are faster and cheaper, instead of only producing more slow UI tests. [Agent architecture and orchestration; AI evaluation and quality] (https://farhanbaluch.com/work/copilot-agent-program.md) - 2026-02-09: **Caught 9 of 9 injected bugs with risk-based test design.** Ran a standard test set and an agent-assisted, risk-based set against the same feature with seeded bugs. The standard set caught 1. AI angle: Hard evidence that how agents are directed matters more than how many tests they write. [AI evaluation and quality; Agent architecture and orchestration] (https://farhanbaluch.com/work/risk-based-test-design.md) - 2026-01-19: **Mapped a legacy codebase into a knowledge graph.** Turned repositories into a Neo4j graph of modules, calls and data flows, so engineers and agents query structure instead of grepping. Code navigation got 70% faster. AI angle: GraphRAG over code. Agents get relationships (who calls what, where data flows) that plain vector search misses. [Knowledge, retrieval and context; Agent architecture and orchestration] (https://farhanbaluch.com/work/legacy-modernization.md) - 2025-12-08: **Finished the report migration ahead of plan.** Crystal Reports moved to JasperReports with multimodal agents that compare old and new report output. The modernization ran 70% faster than planned. AI angle: A vision-capable agent checks the migrated report against the original, which is what made automated migration safe enough to trust. [Agent architecture and orchestration; Solution architecture and delivery] (https://farhanbaluch.com/work/legacy-modernization.md) - 2025-11-10: **Moved Visual Basic modules to Java 21 with spec-driven agents.** Ad-hoc AI coding gave inconsistent results. Each module now gets a written spec first, then agents implement and test against it. AI angle: Spec-driven development makes agent output repeatable. The spec is the contract the agent is held to, and the CI loop lets it correct itself. [Agent architecture and orchestration; AI adoption at enterprise scale] (https://farhanbaluch.com/work/legacy-modernization.md) - 2025-08-18: **Cut manual estimation effort by 80%.** The estimation workflow went live with SME review and an immutable audit trail. Estimators now review and adjust a cited draft instead of starting from a blank sheet. AI angle: Agents do the research and drafting, and a person keeps the decision. Human-in-the-loop plus an audit trail is what let a bank put it into production. [Agent architecture and orchestration; Solution architecture and delivery] (https://farhanbaluch.com/work/estimation-platform.md) - 2025-05-26: **Designed a five-agent estimation supervisor.** Estimates started from long, unstructured requirement documents and varied by estimator. Built a LangGraph supervisor coordinating research and estimation agents. AI angle: A supervisor pattern with typed state and rollback, so each agent has one job and the workflow can recover when one fails. [Agent architecture and orchestration] (https://farhanbaluch.com/work/estimation-platform.md) - 2025-03-03: **Built an LLM crawler for 20+ bank websites.** Selector-based scrapers broke every time a site changed. The crawler lets a model read rendered pages into a fixed schema instead. AI angle: Uses the LLM for what it's good at, reading messy pages like a person, and constrains it with a fixed output schema. [Knowledge, retrieval and context; Agent architecture and orchestration] (https://farhanbaluch.com/work/market-research-agents.md) - 2024-12-09: **Built a realtime voice agent for banking onboarding.** A LiveKit and OpenAI Realtime voice agent that walks new customers through account opening, with handoff to a human when it isn't sure. AI angle: Speech-to-speech agents with a confidence-based handoff. Building one taught me where voice agents fail, which later shaped how I test them. [Agent architecture and orchestration; Solution architecture and delivery] (https://farhanbaluch.com/work/banking-voice-onboarding.md)