The Toolmaker’s Thesis: Confidence is a Function of Control
Most modern software engineering is built on determinism: given state A and function f, you receive output B every time. You write type checks, enforce integration boundaries, and compile with confidence because the execution envelope is declared upfront.
Then came generative AI.
Almost overnight, the industry abandoned decades of systems discipline in favor of prompt chaining. Developers began handing probabilistic models unconstrained access to databases, APIs, and payment rails—hoping an LLM wouldn't drop context, hallucinate a parameter, or mutate production state.
PATHSUM exists because "prompt and pray" is not an engineering paradigm.
If an autonomous workflow issues a quote, commits capital, or mutates a business record, that execution path cannot be left to chance. Our thesis is simple: Confidence is a function of control. In an agentic system, control is the sum of declared constraints—compile-time envelopes, sandboxed runtime permissions, deterministic human gates, and tamper-evident audit receipts.
Built for the Reality of Production
PATHSUM is developing an open-core execution engine and a local-first desktop workbench. All software is pre-release; the planned controls include:
- Compile-Time Envelopes (PathPLANs): Declare routes, validation schemas, token budgets, and execution boundaries before the model touches an environment.
- Runtime Sandboxing (Sealed Tools): Strict capability allowances. Tools execute under the principle of least privilege—never open-ended system access.
- Deterministic Interventions (Human Gates): Enforce hard operational checkpoints before irreversible state changes occur.
- Cryptographic Verification (PathPROOF): Portable, verifiable receipts of every decision, tool call, and state transition.
Control also makes the work useful beyond a single run. Inputs, outputs, human-gate decisions and outcomes stay connected. PathLAB carries that operating evidence into experiments and reusable datasets for training models or capturing company processes.
Behind PATHSUM
My focus on systems didn't start in an academic lab; it started with practical, operational friction.
Early on, I watched a colleague strain to manage a team of 35 people using pen, paper, and ad-hoc spreadsheets. Building software to replace that setup taught me a fundamental reality about how people actually work: people get used to things being the way they are. For anyone to abandon a familiar habit and switch systems, the jump must be easy, and the improvement must be undeniable.
If adopting a tool requires tearing down an entire workflow, people stick with what they know. And if the new system requires constant second-guessing, the improvement isn't real—it's just friction in a different wrapper.
When autonomous AI arrived, that same barrier resurfaced. Teams retreated to manual oversight on critical tasks not because they lacked ambition, but because agentic tools demanded an enormous architectural leap while delivering unpredictable, probabilistic results.
I built PATHSUM and PathLAB Studio directly around those two requirements:
The jump is easy: It is local-first, lightweight, and stack-agnostic. You don't have to overhaul your models, rewrite your codebase, or pipe proprietary data into another third-party cloud dashboard. It runs directly on your machine as an integrated workbench and runtime.
The improvement is undeniable: You move from hoping an agent gets it right to enforcing compile-time boundaries, sandboxed tool execution, and cryptographic audit receipts. When an action executes, you have verifiable proof of how it happened.
Explore PATHSUM
All PATHSUM software is pre-release, including the runtime, CLI, verifier and PathLAB Studio. None is publicly available to download or use yet.
PathLAB Studio is expected around January 2027; release timing is an estimate. Get in touch to ask about the planned workspace or discuss your automation needs.
Get in Touch: Reach me directly at john@pathsum.ca.