Toward an Agentic OS
A workflow that runs 10 steps at 90% per-step reliability succeeds 35% of the time. At 20 steps, that's 12%. The exponential degradation of workflow performance with complexity means that even if models improve to 99% reliability, degradation of long multi-step workflows is inevitable.
What Agentic Reliability Actually Means
As we deploy agents into PHIN OS, trust and verifiability are our central concerns.
Agents break the paradigm of deterministic software in two ways. First, their generative architecture is indeterminate by nature and more importantly the goals we are tasking them with are open-ended.
To achieve reliability, we need to make sure that the agentic harness preserves as much determinism as possible. Central to that aim is to minimize the length of individual workflows through a paradigm of component reuse.
In PHIN OS, this takes three concrete forms:
- Task/Experiment Retrieval. Rather than letting agents plan experiments by default, we expose a library of validated experiments that agents use as primitives. These are validated and calculate specific materials properties. When an agent retrieves a validated experiment, it eliminates all generation failures entirely. The only task the agent has to do is to fill in known parameters for the template.
- Experiment Generation. Experiments for materials properties that cannot be retrieved are generated. Experiments themselves are composed of components. At the base are atomic components that are foundational operations. On top of these are experiment components that are their own building blocks. As agents build workflows out of components, workflows themselves become retrievable components.
- Structured Output Contracts. Each component defines its inputs and outputs explicitly. This means agents can verify that their compositions are well-formed before execution, catching type mismatches and missing fields at planning time rather than at runtime. The agent's job is orchestration, not construction.
To ensure a high standard of reliability, we treat human oversight as a feature, not a liability. Every generated experiment is marked as "Agent Generated". Before it can be widely reused, our team reviews its accuracy and marks it as "Human Reviewed". As agents build on top of these verified components, the roll-out length of individual agent generation decreases, exponentially increasing reliability.
PHIN’s Agentic OS Today
This philosophy allows us to build a platform where agents handle increasingly complex workflows autonomously without sacrificing reliability or quality. This frees users to focus on decisions that genuinely require human judgment. Stage-gating deployments ensures that agents are capable of working autonomously on difficult tasks that increase in complexity because they have a validated component library that is growing.
To see this plan in action, we recently tasked our agents to create 25 new experiment workflows that our existing component library could create. These cover structural, thermal, defect, surface, mixture, and interfacial properties. They are human verified and ready to use with materials across the periodic table. Interested in trying them out? Schedule a demo today.