THESIS 001 REASONING INTO RUNTIME

An LLM should not have to reason through the same task twice.

Agents discover how to complete new work one decision at a time. Applied Runtime turns the parts they have already figured out into secure graph programs that can run 80 to 90 times more efficiently.

READ THE THESIS OUR FIRST PRODUCT: ANSWERTHIS

A complex reasoning trace being compiled into a verified graph runtime shared across permissioned scopes

THESIS / A

The output of reasoning should not only be an answer. When the work is stable, the output should also be a program that never needs to be discovered from scratch again.

OBSERVATION 0001
REASONING IS THE DISCOVERY PHASE

01 / THE THESIS

Reason once. Compile what repeats.

An LLM agent writes the procedure for a task while it performs it. It interprets the request, decides what to do next, selects a tool, transforms data, checks the result, and recovers when a step fails. Systems such as HERMES Agent demonstrate how far this approach can go. For new and ambiguous work, that flexibility is the point.

The problem begins after the system has already learned the procedure. Most production agents still reconstruct every step for every user, spending tokens and time to rediscover a path that is already visible in prior traces. Applied Runtime treats a successful trace as a candidate program. Stable work is compiled into a graph runtime. Novel work stays agentic. One pipeline chooses the right path for each request.

Reasoning discovers the program

An agent trace contains a candidate procedure, not just an answer.

Repetition earns compilation

A workflow is created only when the evidence says the task is stable enough.

Graphs preserve control

Typed inputs, explicit branches, bounded tools, tests, and rollback replace hidden improvisation.

Reuse follows authority

A runtime can be private, organizational, or shared across organizations only when its data and permissions allow it.

02 / THE SYSTEM

From reasoning trace to reusable runtime.

Reason through the new task

The harness decomposes the goal, chooses tools, handles exceptions, verifies intermediate results, and reaches a complete answer. This is where flexible reasoning is valuable.

Capture the procedure

The trace is more than a log. It is a record of decisions, tool contracts, data dependencies, checks, retries, and the conditions under which the task succeeded.

Find the stable subtask

Across related requests, we identify the region whose inputs and outputs are stable enough to separate from open-ended reasoning. The boundary may be a tool, a workflow, or an entire task.

Compile a graph runtime

The proven procedure becomes a typed graph with explicit branches, permissions, tests, and recovery paths. Known work can now execute directly instead of being reconstructed token by token.

Promote with evidence

A graph is promoted only after it clears correctness, security, and scope checks. Every version remains observable, reversible, and linked to the traces that justified it.

Execute, measure, and fall back

Matching work runs through the graph. New or uncertain cases return to the agent. Each execution produces evidence that can strengthen, narrow, or retire the runtime.

PERFORMANCE / MEASURED TARGET

Up to 80 to 90 times more efficient on work the system has already learned.

NOVEL TASK / AGENT REASONSKNOWN TASK / GRAPH EXECUTESUNCERTAIN INPUT / SAFE FALLBACKPROVEN PATH / VERSIONED RUNTIMEPRIVATE DATA / USER OR ORG SCOPESHARED LOGIC / EXPLICIT PROMOTION

03 / RESEARCH PROGRAMS

The hard part is deciding what deserves to become infrastructure.

R.01ACTIVE

Task boundaries

Where does a reusable task begin and end inside a long agent trace? We study how to isolate stable subgraphs with explicit inputs, outputs, and failure conditions.

R.02ACTIVE

Compilation thresholds

When has a task been solved often enough to compile? We study evidence, repetition, variance, and evaluation thresholds that justify promotion into a runtime.

R.03ACTIVE

Secure graph execution

How can a compiled graph inherit least privilege, sandbox boundaries, provenance, approvals, and rollback from the agent that discovered it?

R.04ONGOING

Organizational memory

How should a proven runtime move from one user to a team, an organization, or a trusted network of organizations without leaking data or authority?

A useful compiler must know when not to compile. We publish the questions and findings that can move the field forward while keeping user data, production methods, and security-sensitive details private.

PRODUCT 01AnswerThis

Precise scientific AI workflows for real research tasks.

PRODUCT 01 / IN PRODUCTION

AnswerThis

The first product from the founders of Applied Runtime, and the place where this thesis became obvious.

AnswerThis runs scientific research tasks from search and evidence synthesis through citation-backed writing. While building it, we spent roughly half our time reviewing traces, recognizing repeated procedures, and asking coding agents to turn those procedures into tools. The manual work was valuable because it improved speed, cost, and reliability. It was also a compiler being operated by hand.

VISIT ANSWERTHIS

ORIGIN / 2026

We were already doing this by hand.

Applied Runtime asks how to automate the whole transition: where to draw the task boundary, when the evidence is strong enough to create a runtime, how to execute it securely, and how to share the right parts across an organization or a trusted set of organizations.