TRUST-A

Open source · MIT · v0.1

A practical reliability framework for AI agents

What must be true before we can trust this workflow to act?

Make sources, hard rules, uncertainty, permissions, evidence and autonomy explicit before an agent acts.

Explore TRUST-A on GitHub Run the 5-minute review
TRUST-AReliability system
01Truth
02Rules
03Uncertainty
04Scope
05Traceability
06Autonomy
01TTruthWhat proves the facts the agent is using?02RRulesWhich constraints must remain deterministic?03UUncertaintyWhat happens when the system cannot prove something?04SScopeWhat can the agent read, decide, modify or trigger?05TTraceabilityCan the decision and action be reconstructed?06AAutonomyHow much execution authority is justified?
01THE RELIABILITY PROBLEM

The prompt is only one part of the system.

Once an agent reads multiple sources, calls tools, modifies a system or acts without validation, reliability becomes an engineering question.

Q1Which source is authoritative when information conflicts?

Q2Which rules remain non-negotiable outside model reasoning?

Q3Does missing or stale evidence block the action?

Q4What can the agent actually access or change?

Q5Can the decision and its external effects be reconstructed?

Q6How much autonomy does consequence and reversibility justify?

02THE SIX DIMENSIONS

Six questions before more responsibility.

The dimensions are connected. Weak truth should reduce autonomy. Strong rules and narrow scope can limit the consequence of model error.

T01 / 06

Truth

What proves the facts the agent is using?

Define authoritative sources, provenance, freshness and what happens when sources disagree.

R02 / 06

Rules

Which constraints must remain deterministic?

Keep critical invariants in code, policy, configuration or a validation layer the model cannot bypass.

U03 / 06

Uncertainty

What happens when the system cannot prove something?

Preserve missing, stale and conflicting evidence as explicit states, with safe stop conditions.

S04 / 06

Scope

What can the agent read, decide, modify or trigger?

Bound tools and external effects to the minimum capability required for the agent's role.

T05 / 06

Traceability

Can the decision and action be reconstructed?

Retain the sources, rules, unknowns, versions, approvals, tool outcomes and actual actions that mattered.

A06 / 06

Autonomy

How much execution authority is justified?

Match autonomy to consequence, reversibility, evidence quality and the strength of the other five dimensions.

03FROM PROMPT TO RELIABLE SYSTEM

Models are probabilistic.Reliability is built around them.

TRUST-A makes the sources, rules, uncertainty, scope, traceability and autonomy around the model explicit.

Common abstraction

Prompt → Model → Action

A capable model does not make the surrounding workflow reliable by itself.

  1. 01Authoritative sources
  2. 02Validation and freshness
  3. 03Deterministic rules
  4. 04Agent reasoning
  5. 05Action preconditions
  6. 06Bounded capability
  7. 07Decision trace

Reliable agentic workflow= trusted inputs + deterministic guardrails + explicit uncertainty + bounded capabilities + reconstructible decisions + proportional autonomy

04PRACTICAL ASSETS

Use it now. Keep GitHub canonical.

The framework, templates and Skill are open source. Start with the smallest artifact that matches the consequence of your workflow.

01Full frameworkThe canonical explanation of the six dimensions and how they work together.Read the framework ↗02TRUST-A LiteA five-minute reliability review to run before shipping or increasing autonomy.Run the review ↗03Agent ContractA reusable template for sources, hard rules, stop conditions, capabilities, approvals and trace.Use the template ↗04TRUST-A Audit SkillAn installable AI Skill for auditing agents, workflows, prompts and repository instructions.Open the skill ↗
05AUTONOMY IS A DESIGN DECISION

More autonomy is not automatically better.

Higher consequence and lower reversibility require stronger evidence, tighter deterministic controls and a smaller blast radius.

Reversibility is often a better autonomy signal than raw action size.

  1. 01
    READ_ONLY

    Inspect and summarize.

  2. 02
    RECOMMEND

    Propose decisions or actions without executing them.

  3. 03
    ACT_WITH_APPROVAL

    Execute only after an explicit human approval gate.

  4. 04
    ACT_AUTONOMOUSLY

    Execute within an intentionally bounded, observable and reversible scope.

06WHEN TO USE IT
APPLY TRUST-A

When actions have real effects.

  • Tools or connected systems
  • Multiple or changing sources
  • Repository, database, SaaS or production writes
  • Scheduled or autonomous execution
  • Recommendations that trigger real-world actions
KEEP IT PROPORTIONAL

Skip the full review for low impact work.

  • A simple rewrite
  • An isolated brainstorming session
  • A low-impact assistant interaction

TRUST-A is intentionally lightweight. Do not manufacture governance overhead.

OPEN SOURCE · MIT · v0.1

Make the system around the model explicit.

Explore TRUST-A on GitHub