halftrace 0.0.1

Diagnostics for agent loops. Measure how fast each capability decays over long trajectories.

halftrace

Diagnostics for agent loops. Measure how fast each capability decays over long trajectories.

Agents fail in four distinct ways as trajectories get longer: they forget state, drift from instructions, repeat tool calls, and terminate prematurely. Each has a different halftrace - the trajectory length at which the behaviour is half-degraded. Existing eval frameworks measure task success at one point and miss all four curves. halftrace measures them directly.

The halftrace concept

A halftrace is the trajectory length, in tool calls, at which a given agent capability is degraded by 50% relative to its baseline at low N.

Different capabilities decay at different rates. A model's instruction-following might have a halftrace of 30 tool calls while its state recall has a halftrace of 150, meaning instruction adherence falls off five times faster than memory.

The library measures four halftraces per (agent, model, task):

Probe	What it measures
state_amnesia	Retention of facts introduced earlier in the trajectory
instruction_decay	Adherence to system-prompt rules over time
tool_repetition	Avoidance of re-calling tools with the same arguments
premature_termination	Completing the task before declaring done
narration_substitution	Emitting tool calls rather than describing them

What this is

A measurement library for agent trajectories. You bring the agent. halftrace instruments the trajectory and tells you which capability decays first.

The library was built to answer one question: when an agent fails on a long-horizon task, which failure mode caused it? The answer matters because the failure modes have completely different fixes: better prompting, better tools, better memory, or a different model, and you can't choose without measuring them separately.

What this isn't

• Not a benchmark. No leaderboard, no canonical task set. You define the tasks.
• Not an eval framework. If you want grading, scoring rubrics, or production observability, use Inspect, Braintrust, or Langsmith. halftrace is a diagnostic instrument that sits alongside them.
• Not an agent framework. It doesn't build agents. It measures agents you've already built.

Install

Once 0.1.0 ships:

pip install halftrace

Optional extras:

pip install "halftrace[anthropic]"
pip install "halftrace[openai]"
pip install "halftrace[all]"

Requires Python 3.11+.

License

MIT.