Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.phrony.com/llms.txt

Use this file to discover all available pages before exploring further.

Workflows vs. agents

Not every problem requires an agent. In many cases, what is described as an “agent” is in fact automation — a system where both the objective and the execution steps are known in advance. A workflow is a program with a known path. It encodes decisions ahead of time and produces a predictable outcome. The inclusion of an LLM does not change this: if both the goal and the method are already specified, the system is not acting autonomously; it is following instructions. An agent is a system that determines its next action during execution, operating under uncertainty in pursuit of a goal. Instead of prescribing a fixed path, the system is given an environment, a set of capabilities, and an objective. The sequence of actions emerges at runtime. Agents operate in a continuous loop of perception → reasoning → action → adaptation.

When Phrony fits

Phrony is not intended for every problem. If a system can be expressed as a fixed sequence of steps, it should be. Deterministic workflows are easier to reason about, cheaper to operate, and significantly more predictable. Phrony becomes relevant when the path to the outcome is not known in advance and cannot be reasonably encoded upfront. This typically occurs in four scenarios:

1. Uncertainty in decision-making

The number of possible paths grows combinatorially, making upfront encoding impractical. Not all relevant possibilities can be known in advance. Each new conditional branch adds complexity and maintenance cost until the system becomes a dense graph of edge cases.

2. Dynamic information discovery

Workflows assume all required inputs are available at the start of execution. In many real-world problems this is not true. The system must decide what information is missing and how to obtain it — by querying external systems, asking follow-up questions, or exploring alternative sources — before it can proceed.

3. Non-trivial tool interaction

In workflows, tool usage is predefined: which tool to call, in what order, and under which conditions. This breaks down when tool selection depends on context that evolves during execution. The system must decide not only how to use tools, but which tools to use and why.

4. Iterative reasoning

Some problems require multiple passes to reach a conclusion. The system must form hypotheses, test them, revise assumptions, and converge over time. Intermediate results influence future decisions, creating a feedback loop that cannot be modeled as a static flow.

Tradeoffs

Adopting Phrony means accepting a different execution model. The benefits of flexibility and adaptability come with clear trade-offs.

Predictability

Workflows are deterministic by design. Given the same inputs, they tend to produce the same outputs. Agentic systems powered by LLMs do not offer the same guarantees. Behavior can vary across runs due to differences in context, intermediate results, or model reasoning. Phrony constrains but does not eliminate this variability. By governing how decisions are executed, how tools are accessed, and how state evolves, it reduces the surface area of unpredictability and makes system behavior more consistent.

Debugging

In a workflow, failures trace to a specific step in a predefined path. In an agentic system, the path itself is constructed during execution. Phrony addresses this by making execution observable: capturing the reasoning flow, recording all tool interactions, and enabling inspection of how and why decisions were made. While debugging remains more complex than in deterministic systems, it becomes tractable.

Reproducibility

Agentic systems may not produce identical outcomes across runs, particularly when external data or probabilistic reasoning is involved. Phrony mitigates this through structured logging, traceability, and replay mechanisms — allowing executions to be analyzed and, to a reasonable extent, reconstructed.

Operational complexity

Delegating decisions to runtime introduces inherent complexity. Without proper controls, agentic systems can behave unpredictably, access unintended resources, or produce unsafe outcomes. Phrony’s guardrails — operation allowlists, approval requirements, anomaly detection, and execution limits — exist to close this gap.

Design shift

Adopting Phrony requires a shift in how systems are designed:
Traditional workflowPhrony (agentic)
Encode decisions upfrontDefine constraints and boundaries
Fixed execution pathExecution path emerges at runtime
Deterministic outputGoverned, observable output
Easy to debug (predefined steps)Requires execution tracing
PredictableFlexible
The guiding principle: Phrony brings agentic systems closer to the reliability of traditional workflows, without removing their flexibility.