What are the different kinds of agent behavior drift?

One 2026 taxonomy splits agent behavior drift into three kinds: semantic drift, a gradual departure from the original task intent; behavioral drift, the agent settling into strategies or tool use nobody specified; and coordination drift, multi-agent agreement breaking down over a long interaction. Semantic drift shows up as the output wandering from what was asked, while coordination drift only shows once you compare what agents agreed on earlier against what they agree on now.

The paper is a single-author, simulation-based study, not a measurement on production agents, and its own simulations found semantic drift appearing in close to half of workflows by 600 interactions.

Tessary’s own behavior_drift classifier watches one of the three: it learns a call site’s normal sequence of steps and flags a trace that departs from it, which is the shape behavioral drift takes. It has nothing to say about whether a single agent’s answers have wandered from the original ask without changing which steps it took, or whether two agents in a handoff still agree; an LLM judge is the check built for that.

sources

"Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions" (arXiv:2601.04170) fetched 06 oct 26

keep reading

More on this.

What does a behavior drift baseline catch, and what does it miss?It catches a new or reordered step almost every time and misses a real share of the times an agent quietly stops doing something it used to do.answer →What is Tessary's behavior_drift classifier?It learns an agent's normal steps and their order from that call site's own history, then flags traces that depart from it. It doesn't judge whether an answer was right.answer →Why did my agent stop doing a step it always used to do?Something changed upstream, a prompt, a tool, or a logic branch, and Tessary's behavior drift classifier is built to flag that kind of omission.answer →Everything on tessary behavior driftThe definition, and every question answered under tessary behavior drift.concept →

Two ways to run Tessary.

Tessary is an open-source agent reliability platform. Cloud and self-hosted run the same workflow on the OpenTelemetry traces your agent already emits.

Tessary Cloud

We host it for you. Send your first trace with nothing to deploy and no model key.

what's includedper organization

traces: 10,000 per calendar month
stored trace data: 1 GB
retention: 30 days
model credit: $10, one-time, for triage and root-cause analysis
credit card: not required

start for free

Self-hosted Tessary

Run the open-source code on your own infrastructure with one command. Add your own model key for triage and root-cause analysis.

set up with a coding agentinstall manually

Self-host Tessary for me by following https://github.com/tessaryai/tessary/blob/main/setup.md

docker compose -f oci://docker.io/tessaryai/tessary:compose up -d -y