Raman Shrivastava
← Teaching

Track 02

Agentic Distributed Systems

Your agent already is a distributed system. Here is the theory that explains why it breaks.

9 modules · 0 lessons published

Module 01soonThe Agent as a NodeYour agent is a JSON-RPC node that hallucinates. Most of distributed systems still applies.
  • Tool calls as RPC: HTTP, MCP, A2A as the agent transport
  • Agent identity, addressing, and capability declaration
  • Idempotency keys and retry semantics for tool calls
Module 02soonTime, Order, and CausalityAsync tools, retries, and nested agents have no global clock. Logical time still works.
  • Why wall-clock time fails for agent traces
  • Lamport and vector clocks applied to agent flows
  • Causality as the basis for debugging multi-agent loops
Module 03soonDiscovery, Routing, and GossipMCP registries, A2A directories, and model routers are service discovery — same patterns, new failures.
  • MCP server registries: agent service discovery
  • A2A and emerging agent directory protocols
  • Model routing as cost-and-capability-aware load balancing
Module 04soonState and Consensus Across AgentsTwo agents asked the same question give two answers. Here is when consensus matters and when it does not.
  • Why agents disagree even with identical prompts
  • Shared scratchpads, CRDTs, and agent memory consistency
  • Leader election among agent ensembles
Module 05soonCoordinating Long-Running WorkSagas, 2PC, orchestrator-worker — for tool calls that must commit together and agent flows that span hours.
  • Sagas for multi-step agent flows with compensation
  • Two-phase commit across atomic tool sequences
  • Orchestrator-worker patterns for parallel fanout
Module 06soonScaling Agent SystemsPrompt caches, semantic caches, model gateways, vector indexes, work queues — the agent CDN stack.
  • Prompt caching as distributed caching, with new failure modes
  • Semantic versus exact caches and cache invalidation
  • Agent gateways and proxies (LiteLLM, OpenRouter, internal gateways)
Module 07soonAgent Memory & StorageAppend-only memory logs, artifact stores, and how to migrate a memory schema without amnesia.
  • Append-only memory logs and replay
  • Artifact stores for agent outputs
  • Compaction strategies for long-running agents
Module 08soonOperating Agent SystemsTracing 100 agents at 3am, prompt injection as a trust boundary, schema rollouts that do not break running agents.
  • Distributed tracing for agents (LangSmith, OpenTelemetry GenAI)
  • Prompt injection as a security boundary problem
  • Sandbox isolation and per-agent IAM
Module 09soonSharding, Fanout, and Byzantine AgentsParallel agent fleets, map-reduce fanout, and what to do when sub-agents lie.
  • Work partitioning across agent fleets
  • Map-reduce style fanout (orchestrator/worker)
  • Hot-shard problems with cost-asymmetric agents