Skip to main content

cmd+ctrl

Resources for engineering an AI company.

Technical guides, field notes and practical explanations from the team of Trampoline.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

Governance and security

Bound Context and Writes in Enterprise AI Agents

An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.

Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.

Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.

Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.

All articles

An incoming request passes through a single contextual decision point, then follows a defined sequence of execution steps.
Building AI systems

When AI Workflows Need Judgment vs. Deterministic Steps

Decompose work before choosing how to run it. Use a model for context-dependent decisions and deterministic execution for actions whose path is already defined.

Read article
A durable external repository remains connected as interchangeable task sandboxes are repaired or reassigned around it.
Building AI systems

Treat Sandboxes as Resources, Not an Agent’s Home

A fixed VM for every agent can turn machine upkeep into an agent-platform problem. Separate durable agent work from task execution, then choose capacity that fits the task.

Read article
An abstract agent receives only policy-filtered records and sends a proposed change to a separate human approval point before a protected system is updated.
Governance and security

Bound Context and Writes in Enterprise AI Agents

Enforce permissions before material enters model context, then turn consequential writes into reviewable plans that require explicit human approval.

Read article
Two complete model–harness assemblies balance on a scale above equal stacks of budget tokens, emphasizing fair matched-expenditure comparison.
Evaluation and reliability

How to Evaluate a Long-Horizon Agent Harness

A decision-grade evaluation compares complete model–harness configurations fairly, identifies where extra inference yields marginal returns, and verifies that success occurred in a valid environment.

Read article