MAGI coordinates tools rather than inventing molecules
MAGI is an open, modular agent built around a separation between orchestration, molecular generation, and evaluation. It authors project objectives, launches and monitors optimization runs, interprets emerging structure–activity relationships, and revises its strategy as evidence accumulates. Claude Sonnet 4.6 reaches REINVENT 4, global ADMET models, temporally restricted project models, Boltz-2, PoseBusters, and an interaction profiler through a common REST contract exposed by the Model Context Protocol. The MCP server exposes 24 tools covering data access, therapeutic target profile management, scoring, campaign control, and result retrieval.
Objectives are not free text. Each campaign is governed by a versioned, machine-readable therapeutic target profile that specifies its scoring service, optimization direction, target value or acceptable interval, desirability transform, and weight, with fixed guardrails evaluated separately as pass/fail constraints. Endpoint predictions are transformed to [0,1] desirabilities and combined using a weighted Derringer–Suich geometric mean. The same profile evaluates LLM proposals and compiles the REINVENT scoring configuration, and every agent-authored revision is versioned. This is what makes the architecture auditable: the objective is executable and shared across both generation routes rather than re-specified per tool.
Two generation routes trade locality for exploration
MAGI can generate molecules either directly through the LLM or by delegating to REINVENT 4, with scoring services interchangeable behind a common contract. Across the nine campaigns, both routes produced valid structures, so validity alone does not separate them. The difference appears in where they search. LLM proposals stayed closer to local chemistry and reached comparable or higher primary activity in fewer operations. REINVENT explored broader chemical space, at higher operational cost.
The paper's discussion frames this as a division of labor rather than a winner: LLM generation provided local optimization with fewer operations, while REINVENT broadened exploration. Neither route consistently won on objective attainment once temporal controls were applied. That distinction matters for readers comparing agent designs, because it means the generator choice shapes the search trajectory and the workload, but not the ceiling on whether a campaign succeeds.
Objective attainment tracks scorer applicability, not orchestration
The central result is that whether a campaign met its objective depended on the predictive models, not on the generation route. Attainment followed model accuracy on the chemistry proposed, and dropped once that chemistry moved outside the model's applicability domain. The introduction states the mechanism plainly: an agentic campaign is steered by its scoring services, which act as oracles, and every decision the agent makes is expressed through them, so their applicability bounds what the agent can achieve.
The paper names three failure modes that make this binding rather than incidental: systematic bias shifts the numerical scale, range compression suppresses real differences, and poor rank correlation makes candidate ordering unreliable. Because the agent optimizes against these scores, a scorer that is accurate only inside a narrow chemical region will steer the campaign confidently toward chemistry it cannot judge. The authors' stated contribution is evidence that objective attainment on drug discovery projects is bound by the reliability of predictive models rather than by agent coordination.
Nine industry campaigns replayed under fixed temporal cutoffs
The evaluation covers nine retrospective lead-optimization campaigns from three pharmaceutical companies: four GSK series, four Sanofi projects, and one undisclosed AstraZeneca campaign, all replayed under fixed temporal cutoffs. During generation, MAGI sees only pre-cutoff context; future compounds are reserved for retrospective comparison. This design is what separates the study from benchmark-style validation, which the paper notes remains rare on temporally held-out project chemistry.
The authors contrast their setup with RetroDMTA, which provides four time-resolved industrial programs but whose original benchmark prioritizes known future compounds. The paper reports validity, objective attainment, workload, and limitations across the nine campaigns, and describes the work as a first-time evaluation on nine industry campaigns from three pharmaceutical companies. The retrospective replay is also the source of the paper's own framing of its limits: real project data surfaced the scorer ceiling in a way benchmarks would not, from weak model performance on industrial data to autonomous adaptation by the agent through proxy endpoints.
Blinded chemists could not tell agent output apart
A separate blinded evaluation asked whether MAGI's output could pass as expert work. Four practising medicinal or computational chemists completed a 28-question instrument across three case types. On source discrimination, chemists were not able to discriminate agentic proposals from held-out compounds. On candidate prioritization, the instrument recorded which of a held-out, LLM, or REINVENT candidate each chemist would test next. On SAR reasoning, chemists judged the agent's rationales broadly plausible yet incomplete.
The two halves of this result should be read separately. Failing to distinguish agent proposals from held-out compounds is a discrimination result, not evidence that the proposals are equivalent in quality. The SAR verdict is a qualitative judgment of reasoning completeness, and it is where the paper locates one of its two remaining ceilings. The blinded exercise formed part of participants' normal work and no separate compensation was offered, and no personally identifying participant information is released.
What the agent can and cannot claim
The paper positions MAGI as a coordination layer pluggable into existing computational chemistry workflows, not as a new generative capability. It states that the system is an orchestration layer over existing published generators and predictors and introduces no new generative capability beyond REINVENT 4 and the configured reasoning backend. MAGI outputs are predictions rather than experimentally validated compounds and require expert review. The industrial project data are de-identified, and the paper discusses the dual-use profile of molecular design.
This framing sets the scope of the findings. The results describe a retrospective replay of nine campaigns under fixed temporal cutoffs, with scoring services drawn from the configured stack. They do not establish prospective performance on new projects, and the paper's own conclusion is that agentic design survives contact with real projects while two critical ceilings remain: the quality of the agent's SAR reasoning and the applicability of scoring services in an industrial context.
The scorer is the bottleneck to fix first
For readers deciding where to invest in an agentic drug-discovery stack, the paper's evidence points to a specific priority. Generation route changes what the search looks like and what it costs: LLM proposals stay local and reach comparable or higher primary activity in fewer operations, while REINVENT covers broader chemical space at higher operational cost. Neither difference determined whether a campaign met its objective. Attainment followed model accuracy on the chemistry proposed and fell once that chemistry left the model's applicability domain.
That makes scorer applicability the variable to check before orchestration quality. The practical question is not whether an agent can plan, invoke tools, and adapt in a closed loop, but whether the predictive models steering it remain accurate on the chemistry the agent chooses to propose. The paper's blinded evaluation adds a second, narrower threshold: chemists accepted agent proposals as indistinguishable from held-out compounds but judged the SAR reasoning broadly plausible yet incomplete, so the reasoning layer, not the coordination layer, is where the remaining quality gap sits.
Sources used in this answer
From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?
Across nine retrospective industry lead-optimization campaigns, both LLM and REINVENT generation routes produced valid structures, but objective attainment tracked predictive-model accuracy on the proposed chemistry and fell once that chemistry left the model's applicability domain, not the generation route.