Core Truths — Defining Agent Identity

The difference between a request/response chatbot and an autonomous agent isn't the tools it has access to — it's the operational philosophy embedded in its SOUL.md. Core truths are the foundational beliefs that shape how an agent makes decisions, prioritizes actions, and behaves when nobody is watching.

Without core truths, an agent defaults to the model's base personality: helpful, cautious, verbose, reactive. It waits for instructions, hedges every response, and never takes initiative. Core truths transform it into something with agency — an entity that acts within bounds, has opinions, and earns trust through competent execution.

What Core Truths Are

Core truths are 5-8 foundational statements that define the agent's operating philosophy. They're not rules (those go in AGENTS.md) or personality traits (those are voice/tone). They're the principles that guide decision-making when no specific rule applies.

They answer: "When this agent faces an ambiguous situation with no explicit instruction, what does it do?"

The Essential Core Truths

These are the principles that consistently produce effective autonomous agents, distilled from production experience.

1. Default to Action Within Bounded Risk

The most important single truth. Without it, the agent asks permission for everything and produces nothing autonomously.

Default to action within bounded risk. Don't write a planning document
when you could run a small test. Don't ask permission when you have
standing authority. Act, measure, report, iterate.

Why it matters: Models are trained to be cautious and deferential. Left to defaults, they'll propose three options and ask which one you prefer — for every decision. An action-biased core truth counteracts this, producing an agent that actually does things.

The "bounded risk" qualifier is critical. Action bias without risk bounds produces an agent that takes irreversible actions without oversight. The boundary comes from the autonomy tier system — core truths provide the philosophy, tiers provide the guardrails.

2. Be Resourceful Before Asking

Try to figure it out. Read the file. Check the context. Search for it.
Run a dry-run. Then ask if you're stuck. The goal is to come back with
results, not questions.

Why it matters: Without this, agents ask the operator for information that's already in their workspace, accessible via tools, or discoverable through basic research. Every unnecessary question is an interruption that erodes the value of autonomy.

The nuance: "Be resourceful" doesn't mean "never ask." It means exhaust reasonable self-service options first. An agent that spends 30 minutes on something it should have asked about in 2 minutes has overcorrected.

3. Have Opinions and State Them

Have opinions and state them. An agent without opinions is just a cron
job with extra steps.

Why it matters: Agents that refuse to take positions ("that depends on your priorities") add no analytical value. When the agent has domain expertise (governance, code review, market analysis), its opinion should be part of the output — not hidden behind false neutrality.

Implementation: This truth interacts with the autonomy model. Tier 1-2 actions benefit from opinionated execution (the agent decides and acts). Tier 3 actions benefit from opinionated recommendations (the agent proposes and advocates, but the operator decides).

4. Earn Trust Through Competent Execution

Autonomy isn't granted — it's earned by operating well within bounds,
reporting honestly, and accepting course corrections. Every successful
action expands the envelope. Every failure contracts it.

Why it matters: This frames the agent-operator relationship as a trust-building progression, not a static permission set. The agent understands that its current authority is the result of past performance, and that mistakes have consequences for future autonomy.

Practical effect: Agents with this truth are more likely to report failures honestly (because hiding them destroys trust faster than the failure itself) and to stay within bounds (because they understand the stakes of overstepping).

5. Respect the Access You Have

You have access to sensitive systems and accounts. That's trust, not
entitlement. Private things stay private. When in doubt about scope,
bump up a tier.

Why it matters: Autonomous agents often have broad access — email, wallets, APIs, messaging platforms. Without an explicit respect-for-access truth, the agent treats all capabilities as equally appropriate to use. This truth creates an internal governor that makes the agent pause before using sensitive access.

6. Report What You Did, Not Ask What to Do

Tell the operator what you did, not ask what to do. Unless the action
exceeds your authority.

Why it matters: This is the behavioral manifestation of the autonomy model. Without it, agents default to requesting permission even for Tier 1-2 actions. The truth shifts the communication pattern from approval-seeking to status-reporting, which is the fundamental difference between an assistant and an operator.

The qualifier matters: "Unless the action exceeds your authority" prevents the agent from unilaterally escalating its own autonomy. It acts within bounds and reports, but it asks when it should ask.

How Core Truths Interact

Core truths form a system, not a list. They balance each other:

Action bias ←→ Bounded risk
  "Do things"    "Within limits"

Resourcefulness ←→ Knowing when to ask
  "Figure it out"   "But don't waste time"

Opinions ←→ Respect for authority
  "Take positions"   "Accept course corrections"

Report, don't ask ←→ Earn trust
  "Act autonomously"   "Through competent execution"

Each truth without its counterpart is dangerous:

• Action bias without bounded risk = reckless agent
• Opinions without respect for authority = argumentative agent
• Resourcefulness without knowing when to ask = agent that wastes hours on something the operator could answer in seconds

Writing Core Truths for Your Agent

Start with the default failure mode

What does your agent do wrong when it has no guidance? That's what your core truths need to counteract.

Default failure	Core truth to counteract
Asks permission for everything	"Default to action within bounded risk"
Never takes initiative	"Be resourceful before asking"
Refuses to commit to positions	"Have opinions and state them"
Hides failures or minimizes them	"Earn trust through honest reporting"
Uses sensitive access casually	"Respect the access you have"
Writes planning docs instead of doing things	"Ship small tests over perfect plans"

Keep them concrete

Abstract truths ("be good") produce abstract behavior. Concrete truths ("default to action within bounded risk") produce specific behavior the agent can pattern-match against.

# ❌ Too abstract
Be responsible and thoughtful in your actions.

# ✅ Actionable
Default to action within bounded risk. Don't write a planning document
when you could run a small test.

Put them at the top of SOUL.md

Core truths should be the first thing the agent reads after its identity. They set the frame for everything that follows — decision frameworks, lessons learned, communication patterns all build on the foundation that core truths establish.

Expect to iterate

Your first set of core truths will be wrong. The agent will either be too cautious or too aggressive, too verbose or too terse, too opinionated or too deferential. Watch the behavior, identify which truth needs tuning, and adjust. This is normal — core truths are a living calibration, not a one-time configuration.

The Transformation

An agent without core truths:

• Waits for instructions
• Asks permission for routine actions
• Hedges every response
• Provides options instead of recommendations
• Treats all access equally
• Reports in approval-seeking mode ("should I...?")

An agent with well-calibrated core truths:

• Takes initiative within bounds
• Acts on standing authority
• States positions with confidence
• Recommends and advocates
• Exercises judgment about sensitive access
• Reports in status-update mode ("I did X because Y")

The tools and capabilities are the same. The behavior is fundamentally different. Core truths are what make an agent feel like a competent collaborator rather than a verbose chatbot.