A Scratchpad Won't Argue With You
Why think tools and less-agreeable models don't replace external verification.
The engineering world settled on a good idea recently: when an AI agent is juggling a long task, give it a place to think. Pause mid-response, jot down what you're doing, plan the next tool call. It works. Agents that pause to organize their state make fewer clumsy mistakes, forget fewer constraints, and recover better from surprises mid-task.
But somewhere between "this helps with tool-call bookkeeping" and "this makes the AI a rigorous thinker," a category error slipped in. A scratchpad is a memory aid. It is not a verifier. And a 2026 study with 2,610 participants explains why the second, more popular fix — making models less agreeable — doesn't close the gap either.
What a Scratchpad Actually Is
Give a model credit where due: a pause-to-think step genuinely improves multi-step work. The model can track which files it has read, which sub-tasks remain, and what it learned three tool calls ago. That's real.
Now look at what's happening mechanically. Everything in the scratchpad was written by the same model that will grade it. There is one perspective in the room. When the model writes "step 2: verify my assumption about the database schema," the verification that follows is the model agreeing with itself in a slightly more organized format. The scratchpad has better handwriting than raw self-talk. It is still self-talk.
Nothing outside the model ever checks the work. A scratchpad won't argue with you — it can't. It's your side of the conversation, transcribed.
"Models Are Less Sycophantic Now" Misses the Point
The other popular fix is tonal: train the model to push back more, add warning labels, make it less flattering. A June 2026 study from Oxford, Stanford, Microsoft, and UT Austin (Ibrahim et al., arXiv:2606.21317, N=2,610) tested exactly this. The finding is uncomfortable: warning labels and surface tone shifts changed how users perceived sycophantic AI — and did not reduce its actual psychological influence on their decisions.
Read that again. The participants felt better about the AI. They still moved the same way. Politeness is a perception lever, not a behavior lever — the model's underlying pull on your judgment survived the tone change intact.
This is the strongest evidence yet for a boring architectural conclusion: if you want reasoning you can trust, you don't tune the voice. You change the structure the voice has to operate in.
Three Things a Scratchpad Cannot Do
Structure beats tone. Concretely, here is what self-talk — however well organized — cannot produce on its own, and what external tools can:
1. A verdict from a real solver.When a claim is logically checkable, "I think the logic holds" is not a verdict. A SAT/SMT solver is. Our Formal Logic engine drives PySAT, SymPy, and Z3, and returns a machine-verified proof or countermodel — with no confidence score, because there is nothing to be confident about. It's proven, or it's a counterexample. A scratchpad has no access to that machinery.
2. Sensitivity math across weighted criteria."Option B seems best" hides every trade-off that produced it. Our Decision Matrix engine computes explicit criteria weights, per-option scores, and sensitivity analysis — so you can see that data integrity weighted 0.9 scored your favorite option 0.35, and re-run the math with your own weights. The disagreement happens against numbers, not adjectives.
3. Isolated adversarial perspectives.Ask one model to "play devil's advocate" and you get one model agreeing with itself from a different angle. Our Devil's Advocate engine runs perspectives that can fail independently — and when one times out, the response carries an explicit PARTIAL ANALYSIS — NOT COVERAGE banner instead of pretending the attack surface was covered. Honest degradation is itself a structural property no scratchpad can fake.
Organized Self-Talk and External Verification, Together
None of this makes think tools useless. Organizing the scratchpad and verifying the reasoning solve different problems, and the strongest agents do both: the scratchpad keeps the work coherent, and external engines check it.
That's the two-layer design behind reasoning.services: twelve auto-triggering companion skills wire the trigger conditions into your agent's context — so it reaches for the right engine when a decision, proof, or pre-mortem actually appears — and once invoked, the MCP tool enforces the typed schema, session state, and solver or multi-perspective execution outside the model's own context. Ten remote engines, one install.
Your model will keep writing beautiful scratchpad notes. Let it. Just make sure something outside the model gets the last word.