Monitoring for
SWE-agent sessions

Why SWE-agent sessions need behavioral monitoring

SWE-agent runs autonomous software engineering sessions that can span hundreds of tool calls. When a session stalls, it often does not crash; it repeats structural patterns. The agent may keep trying the same file edits, searches, or commands until it hits the step budget.

The commercial framing should use percentage of agent budget, not tiny per-session dollars. For example, at $25,000/month in agent/API spend, a 5% reviewable exposure scenario is $1,250/month or $15,000/year. Use the planner for your exact assumptions. This is scenario math, not guaranteed savings.

Validated results on SWE-agent trajectories

A Cohen's d of +1.099 is a large effect in this research slice. It supports structural review, not a production claim that CAUM predicts all session failures.

Integration — 3 lines

CAUM attaches alongside SWE-agent as a passive observer. It receives each tool call and result, classifies the behavioral regime, and emits a running health score. The agent's execution is never modified.

# Works with any SWE-agent version
from caum import ZeroTrustAuditor

aud = ZeroTrustAuditor(model_hint="gpt4o", framework_hint="swe-agent")

# Call after every tool execution in the agent loop
for tool, result in swe_agent.run(issue):
    signal = aud.push(tool, result)
    print(signal["regime"])  # EXPLORER / GRIND / STAGNATION / LOOP

receipt = aud.finalize()  # structural tier, evidence grade, hash-linked receipt

What CAUM detects

Each step is assigned a structural health signal from recent action patterns and event shape:

• EXPLORER - diverse structural action patterns.
• GRIND - slow structural progress and limited variety.
• STAGNATION - weak structural movement; review recommended.
• LOOP - repeated action cycles; cost/token exposure should be reviewed.

Session receipt

Each session can produce a structural observation receipt containing:

• Structural health score - overall trace health
• Tier - T1 through T5 structural health band
• Regime distribution - share of steps in each structural state
• Exposure evidence - loop/stagnation/cycle evidence and token/cost exposure
• Review triggers - specific structural anomaly events detected