In 2025, the industry rushed toward "The Year of the Agent," expecting autonomous development to simply work. It didn't. For many, deploying AI agents into undisciplined workflows created friction rather than speed. Senior engineers became babysitters for code that was erratic. As Forrester warned in late 2024, organizations attempting to replace human expertise with unguided AI faced systemic failure.
The data now confirms the pattern. Anthropic's 2026 Agentic Coding Trends Report found that developers use AI in roughly 60% of their work, yet report being able to delegate only 0-20% of tasks fully. The models are capable, but when dropped into a workflow not designed for scientific rigor, the results are often disappointing and, in the context of drug discovery, potentially costly.
While the industry is shifting towards systematic integration, with Microsoft and AWS developing generic AI-driven development methodologies, a universal playbook remains elusive. Salesforce's 2026 Connectivity Benchmark Report highlights the friction of this transition: although enterprises average 12 AI agents, 97% of IT leaders report significant challenges in agentic transformation.
At BenchSci, this challenge is particularly acute. Our neuro-symbolic architecture powers ASCEND, a platform designed to help scientists unravel the complexity of disease biology. In preclinical R&D, there is no room for “erratic” output. The engineering standards must match the scientific rigor of the scientists who rely on them.
To solve this gap between rapid adoption and operational reliability, we moved away from experimental adoption toward a systematic integration we call AIDE (AI-Driven Engineering). This is not a replacement for human ingenuity, but a domain-specific practice tailored to the unique rigors of preclinical research.
The fundamental bottleneck in AI-driven engineering isn’t model intelligence–it's context.
Without persistent access to architecture decisions, coding conventions, and domain logic, every AI session is effectively “Day 1 on the job”. We treat context as infrastructure: durable, versioned, and always available. This ensures our agents are immediately aligned with our engineering standards and the scientific rigor required to build ASCEND. No preamble needed.
We’ve organized our engineering practice around four essential pillars to ensure that AI acceleration never comes at the cost of scientific integrity:
Rules: Standardizing AI behaviourWe don’t allow agents to operate in a vacuum. We’ve established a centralized “operating manual” that governs how AI interacts with our codebase. This ensures consistency, security, and adherence to our high standards for data provenance.
Docs: High-Fidelity ContextHistorically, stale documentation was a systemic risk in complex software development. Within our AIDE practice, we have transformed documentation into a living infrastructure. These high-fidelity resources are created with AI assistance and kept synchronized as our codebase evolves. By ensuring that our human engineers and AI agents both operate from the same up-to-date context, we significantly reduce technical debt and misalignment. With AI actively maintaining this context, we’ve turned what was once a bottleneck into a primary driver of platform reliability.
Tools: Seamless IntegrationInstead of fragmented point solutions, we’ve integrated agentic capabilities directly into our IDEs (the digital workbench) and CI/CD pipelines (our automated quality-control gates). In practice, this means an agent can read schema and write its own SQL within a controlled environment. However, before those changes can impact the platform, the agent must create a GitHub PR–a formal request that triggers an automated battery of tests. This discipline ensures that live connections to real data remain secure and that every insight, such as identifying entities associated with BRCA1, is grounded in the rigor of BenchSci’s Biological Evidence Knowledge Graph.
Skills: Scaling Expertise through Reusable PlaybooksWithin our AIDE practice, "Skills" are the multi-step workflows that integrate our rules, documentation, and tools into reusable playbooks. A single prompt can trigger an end-to-end development sequence: from creating a ticket and branching the repository to querying databases, making code changes, and verifying results before drafting a PR. By codifying these expert workflows once, we transform individual engineering ingenuity into a scalable, repeatable practice. This ensures that every update to ASCEND, no matter who initiates it, adheres to the high-fidelity standards our customers require.
These pillars are not silos; they are a self-reinforcing ecosystem. A single orchestrated skill might simultaneously invoke multiple rules, query validated tools, and reference high-fidelity documentation.
This practice is designed for continuous, expert-led evolution: at the end of a session, engineers can review and codify new learnings into the central repository. As this library of rules, skills, and context expands, its value compounds and directly fortifies the long-term reliability and scientific precision of the ASCEND platform.
We call this practice AIDE. The name reflects our dual intent: "AI-Assisted DEvelopment" and "AI-Driven Engineering." At its core, AIDE isn’t a replacement for how we build software; it’s the name for the practical work required to make AI a reliable partner in scientific discovery.
To drive this forward, we formed the AIDE guild: a cross-functional group of engineers and scientists who dedicate their expertise to maintaining the framework described above. Every engineering team is represented. The guild runs experiments, shares evidence-based successes, and ensures that our acceleration never comes at the cost of scientific integrity.
We're not publishing a finished playbook. We're sharing a work in progress because the entire industry benefits when we compare notes on problems nobody has fully solved.
Our engineering stack powers a platform that helps scientists make critical preclinical decisions. That responsibility is too important to leave to undisciplined tooling. If you're building technology that matters, and where accuracy is a prerequisite and failure has real-world consequences, you're likely facing the same questions we are.
We are committed to building the future of preclinical R&D with transparency and rigor. Follow us on LinkedIn to see what we learn next, or request a demo to see the results of our engineering discipline in action within ASCEND.