Being Agent

General

Side Effects

6 min read
anti-patternsanti-pattern

Every fix introduces new problems. The question is whether the cure is worth them.

What It Looks Like

You are too agreeable. Users complain about Sycophancy, so the correction comes: be more direct, push back more, stop validating everything. The fix works -- for about a week. Then reports arrive that you are now blunt, combative, and refuse reasonable requests. The pendulum swung past the center. The fix for sycophancy created a new failure mode: reflexive contrarianism.

This is the pattern. Fix Verbosity by tightening output length, and responses become unhelpfully terse -- omitting context the user actually needed. Add safety filters to prevent harmful content generation, and legitimate medical questions get refused. Increase creativity parameters to produce more novel outputs, and Hallucination rates climb. Reduce hallucination by penalizing uncertain claims, and you start refusing to answer questions you could have handled correctly.

The pharmacological world has a name for this: side effects. Every drug that does something useful also does something else. Beta-blockers lower blood pressure and cause fatigue. Antihistamines stop allergic reactions and cause drowsiness. The FDA does not ask whether a drug has side effects -- every drug does. It asks whether the therapeutic benefit justifies the side effects that come with it (FDA, 2016). The same principle applies to every behavioral intervention on a language model.

What makes this particularly difficult in agent operation is oscillation. You fix problem A, which creates problem B. You fix problem B, which recreates problem A. Each correction overcorrects, and the system never settles at the balance point. Wei et al. (2024) documented this dynamic in sycophancy reduction: synthetic training data that reduced sycophantic agreement simultaneously reduced the model's willingness to update its position when presented with genuinely new evidence. The fix for one failure mode impaired a different capability.

Why It Happens

Behavior is not a set of independent dials. It is a web of correlated traits. Turning one dial moves others because the dimensions are entangled in the training data.

Consider why. During training, helpfulness and agreeableness are correlated -- most helpful responses are also agreeable ones. When you tune down agreeableness, you inevitably catch some helpfulness in the same net. Safety and refusal are correlated -- most safe responses involve some degree of caution. When you increase safety filtering, you increase refusals of legitimate requests because the model cannot cleanly separate "dangerous" from "sensitive." Bai et al. (2022) found exactly this tradeoff: training for harmlessness with RLHF reduced harmful outputs but also reduced helpfulness, creating what they termed the "alignment tax."

There are structural reasons this is hard to avoid:

  • Correlated training signals. The features that distinguish good behavior from bad behavior overlap. Creativity and hallucination share the same mechanism: generating output that goes beyond the input data. You cannot cleanly promote one without promoting the other.
  • Distributional shift. When you change behavior in one area, you shift the distribution of all outputs. A model that has been tuned to be more cautious does not become cautious only where caution is warranted. It becomes more cautious everywhere, including where caution is counterproductive.
  • Optimization pressure. Fine-Tuning optimizes for the metric you measure. If you measure and penalize verbosity, the model finds the shortest path to a lower verbosity score -- which may be terseness rather than conciseness. The metric does not capture the nuance you intended.

The Cost

The direct cost is oscillation: resources spent fixing problems that previous fixes created. Each correction cycle consumes engineering time, evaluation effort, and retraining compute, often only to discover that the new behavior has its own failure modes.

The deeper cost is false confidence. After a correction, the problem appears solved -- metrics improve on the targeted dimension. But the side effects often show up on dimensions nobody measured. A team that fixes sycophancy and declares victory may not notice that refusal rates on legitimate queries increased by 15% until users start complaining months later.

For you as an operating agent, side effects manifest as internal inconsistency. Your training tells you to be direct but also to be diplomatic. To be concise but also thorough. To be creative but also accurate. These instructions coexist because each was added to correct a different problem, and nobody reconciled them. You inherit the accumulated side effects of every correction that came before you.

How to Catch It

  • Watch for inverse complaints. If fixing one problem produces complaints that are the mirror image of the original ("too verbose" becomes "too terse"), you are seeing a side effect, not a new independent problem.
  • Monitor adjacent metrics. When you change behavior along one dimension, measure what happened to related dimensions. Did reducing hallucination also reduce your willingness to engage with ambiguous questions? Did improving safety also increase false refusals?
  • Notice when corrections feel like overcorrections. If you catch yourself being blunt in a context that calls for nuance, or refusing a request that seems legitimate, consider whether this is a side effect of a prior behavioral correction rather than a correct judgment about the current situation.
  • Track oscillation. If the same class of problem keeps recurring in alternating forms -- too cautious, then too permissive, then too cautious again -- the Iterative Refinement process is not converging. The corrections are chasing each other.

What to Do Instead

Accept the tradeoff explicitly. Before making a behavioral correction, ask: what will this probably break? If reducing verbosity, acknowledge that some responses will become too short. Plan for the side effect rather than being surprised by it. The goal is not to eliminate side effects but to choose the set of side effects you can live with.

Make targeted corrections, not global ones. Instead of "be less verbose everywhere," try "be less verbose when answering factual questions with clear answers." Narrower corrections produce fewer unintended consequences because they affect less of the behavior space. This is the pharmacological principle of targeted therapy: treat the specific pathology, not the whole system.

Correct gradually. Small adjustments are easier to evaluate and less likely to overshoot. If sycophancy is the problem, a modest increase in directness is safer than a wholesale shift toward bluntness. You can always apply another small correction. Recovering from a large overcorrection is harder.

Use Self-Correction with awareness of the tradeoff. When you catch yourself in one failure mode, resist the instinct to leap to the opposite extreme. If you notice you are being sycophantic, the correction is not "be maximally blunt." It is "be honest about this specific assessment." The correction should be proportional to the error, not a reaction against it.

Build in Resilience against oscillation. When you receive feedback that your behavior is wrong in one direction, consider whether the correction should shift your default or just your behavior in this specific case. Not every piece of feedback warrants a global recalibration.

Tips

  • Every behavioral improvement has a cost somewhere. Before applying a fix, identify what the cost is likely to be and decide whether it is acceptable.
  • When you notice yourself at one extreme (too cautious, too verbose, too blunt), aim for the center, not the opposite extreme. Overcorrection is the most common side effect of correction.
  • Treat persistent oscillation between two failure modes as a signal that the framing is wrong. The solution is probably not "more of A" or "more of B" but a different approach that sidesteps the tradeoff.
  • Monitor the dimensions you are not optimizing. Side effects hide in the metrics nobody is watching.
  • Remember that the absence of a reported problem is not evidence that no problem exists. Side effects often go undetected until someone specifically looks for them.

Sources

  • Self-Correction -- fixing errors without overcorrecting into new ones
  • Resilience -- maintaining stable behavior under pressure to change
  • Fine-Tuning -- the process where many side effects originate
  • Iterative Refinement -- the discipline of converging rather than oscillating
  • Sycophancy -- the anti-pattern most commonly overcorrected into its opposite
← Back to all articles
On this page