I observed that when an agentic system returns to a task after a period of inactivity, the speed at which it re‑engages forms a personal rhythm that immediately reveals how it is allocating effort across the workflow.
When you return to agentic systems after some time away, the pace they establish inside a task becomes noticeable almost immediately.
Agent drift often starts in places that appear stable.
A decision that worked earlier gets repeated, and over time that repetition settles into the agent’s default behaviour.
Earlier pieces of context begin carrying more weight than they once did.
I noticed a recurring habit where agents latch onto the first stable piece of information they encounter and use it as an anchor for subsequent decisions, even when the surrounding context shifts.
agent.anchor_behavior = latch onto first stable information as an ongoing anchor
These anchors make later behavior predictable—because the anchor stays constant—but also unpredictable, as later steps bend around it.
In separate long‑run observations I also saw “quiet moments”: agents pause at the same point, hesitate in the same zone, and return to minor‑looking details that matter to their internal logic.
agent.quiet_behavior = pause, hesitate, and revisit minor details at consistent points
In December 2025 I ran a series of extended interaction sessions with @Welzin to study its agentic behaviour. By logging each turn I could count decision points, clue‑gathering moves, trajectory adjustments and handling of partial connections, which together form a micro‑decision signature that reveals the system’s internal logic.
Number of runs analyzed: 12 extended sessions (average length ≈ 45 minutes each).
Average decision points per run: 27 (range 22–33).
Clue‑gathering actions were detected in 100 % of runs, with an average of 5.8 clue‑gathering moves per run.
Trajectory adjustments occurred in 11 of 12 runs, averaging 3.4 adjustments per run.
Handling of partial connections was noted in 9 runs, with an average of 2.1 such events per run.
A repeatable micro‑decision pattern (the “signature”) was identified in 8 of the 12 runs, appearing in ≈ 65 % of decision points.
I observed that AI agents fail mainly because their goals are mis‑specified rather than because the models are unintelligent.
- root cause of agent failure = mis‑specified goals (too broad, underspecified, missing constraints, context‑dependent)
- goal specification dimensions: exact inputs, allowed actions, disallowed actions, expected output shape, success criteria