conversation dynamics analysis

transition matrix built from 23,262 labeled messages across ~4,656 threads.

transition matrix (row → column)

key findings

healthy patterns

1. NEUTRAL dominates — 77% of neutral messages lead to more neutral. stable equilibrium; the agent is executing without intervention.
2. APPROVAL chains — 38% of approvals lead to more approval. indicates user satisfaction compounds.
3. STEERING recovers — 50% of steering returns to neutral immediately. half of corrections work first try.

doom spiral indicators

• STEERING → STEERING at 30% — nearly a third of corrections require another correction. this is the doom loop.
• only 15 cases of 3+ consecutive steering in entire dataset — rare but distinct failure mode
• FRUSTRATED threads avg 1.7 steering vs RESOLVED at 0.46 — 3.7x higher steering in frustrated sessions
• STUCK thread has 4.0 avg steering — sample size of 1 but fits the pattern

recovery sequences

after STEERING, the most likely paths:

STEERING → NEUTRAL (50%) — immediate recovery ✓
STEERING → STEERING (30%) — correction cascade ⚠
STEERING → APPROVAL (10%) — user confirms fix worked
STEERING → QUESTION (10%) — agent seeks clarification

best recovery signal: when steering leads to approval, the user has validated the correction.

position effects

steering distribution by thread phase:

• early (0-33%): 3.2% of messages are steering
• mid (33-66%): 3.4% steering
• late (66-100%): 3.8% steering

slight uptick late = accumulated frustration or scope drift. early steering more likely about misunderstood intent.

question loops

QUESTION → QUESTION at 39.6% — agent asks, user asks back. not inherently bad (clarification dialogue) but can indicate confusion on both sides.

heuristics

1. 2+ consecutive steering = yellow flag — check if scope was clear
2. STEERING late in thread = possible scope creep — original task may have morphed
3. APPROVAL → NEUTRAL is healthy exit — user approves, agent returns to flow
4. QUESTION chains > 3 = both parties confused — consider reframing the task

thread examples with high steering

high steering doesn’t always mean failure — complex tasks may require more guidance. but UNKNOWN outcomes correlate with higher steering.