Agent vs Tool Distinction
A precise operational distinction articulated by the PeptAI team for autonomous scientific pipelines, sharper than the “agentic AI” marketing usage.
A tool takes an input and returns an output. A sequence goes in, a structure comes back. The decision of what to do with it, whether the structure is good enough, what to run next, whether the candidate advances, stays with the researcher.
An agent uses tools. It evaluates the output against a defined threshold, decides whether the candidate passes or fails, and triggers the next step without waiting for a human to interpret the result.
The shift is not about wrapping a tool in an LLM. It is about moving the threshold and the branching logic out of the human’s head and into a defined contract. Once that contract exists, you can run a fleet of agents in parallel — PeptAI’s setup is one agent per receptor (GLP-1R, KISS1R), each owning the full multi-stage pipeline for its target.
The implication for any computational research pipeline: a .py script is a tool. A .py script that loads a baseline, computes a score, compares against a hardcoded threshold, writes a pass/fail record, and triggers the next stage — that’s an agent, regardless of whether an LLM is involved. The presence of an LLM is orthogonal; what matters is whether the decide-and-trigger step is automated.
Why this matters for STRC
STRC has many tools (AlphaFold3, Boltz2, FEP scripts, electrostatics calculators) but no agent layer. Phase progression is a human reading outputs and deciding “good enough, advance.” The S/A/B/C/D ranking is a hypothesis-level pass/fail framework but doesn’t operate per-phase, per-numerical-threshold. Importing the agent definition would mean: each computational phase declares its threshold up front (e.g., Phase 1d AF3 advances iff pTM ≥ 0.7 AND ipTM ≥ 0.6 AND no clash > 5σ), and the decision becomes mechanical.
This is not the same as “use more LLMs.” It’s the discipline of writing the threshold down before you run the calculation, so the calculation cannot be rationalised post-hoc.
Connections
[source]2026-04-25-bankless-peptai-desci-drug-discovery[applies]STRC Gene Therapy- Nine-Gate Discipline for Computational Drug Discovery — the gate system PeptAI built on this distinction
- Computational Confidence Scores as Epistemic Tools — adjacent epistemics: thresholds tied to specific decisions, not free-floating
[see-also]scientific-research-pipeline