An agent stack inherits its failure modes from the combination you pick

Framework, runtime, coordination pattern, memory, and harness controls get chosen one at a time, but the risks follow the combination: a hub creates a topological bottleneck, multiple agents create cascade amplification and genealogy drift, and durable memory opens a path for context poisoning. The cheapest gap to miss is error handling — five steps at 95% success each finish the whole path about 77% of the time.

Assemble an agent architecture

Choose one option on each architecture axis. The output is a starting contract and risk review. Validate framework support, model behavior, and operational limits in your own environment.

1 Framework

Graph-explicit state machine · Maximum control over state and flow

2 Runtime

You operate the application runtime, network, storage, and deployment controls.

3 Coordination pattern

Medium complexity · Research, complex analysis, or tasks that benefit from specialized sub-tasks.

4 Memory

Working context now + durable recall behind one memory layer.

5 Harness controls pick any
Starting architecture
FrameworkLangGraph — Graph-explicit state machine
RuntimeSelf-managed
CoordinationOrchestrator–Workers — Medium complexity
MemoryHybrid
HarnessObservability, Error handling
Watch for (6)
Cascade amplificationcritical

Fix Require independent re-derivation of load-bearing claims (verification pipeline).

Topological fragilityhigh

Fix Avoid hub bottlenecks; prefer flatter topologies; replicate critical routing roles.

Consensus inertiahigh

Fix Bound the rounds; periodically restart from independent initial conditions.

Genealogy drifthigh

Fix Propagate a genealogy graph with every artifact — who produced it and from what.

Context poisoningcritical

Fix Quarantine untrusted input; per-agent firewalls so bad content can’t enter memory.

Hallucination compoundinghigh

Fix Ground every step with retrieval against authoritative sources; refuse to act unverified.