PRAEVA SYSTEMS RESOURCES

Authority before action.

Research notes, incident analysis, primers, and public doctrine for organizations building agentic systems that can act in consequential environments.

FEATURED SERIES
WHEN AGENTS RUN AMOK #010 · NEW · NOVEMBER 12, 2026 United Said No to AI Agents. The Real Question Is Who Gets to Say Yes. United's restrictions on unauthorized third-party agents raise the question of how to verify consent, platform permission, and delegated transaction authority. Read #010 →
WHEN AGENTS RUN AMOK #009
NEW · October 31, 2026

We Built the Guardrails. Then We Tried to Break Them.

The founder tests stopped unauthorized transactions. But can an agent bypass the gateway, or can the evidence be altered? The next challenge is independent enforcement-path verification.

Read #009 →
When Agents Run Amok series illustration
WHEN AGENTS RUN AMOK #008
NEW · October 16, 2026

We Turned the Orcas Loose

Twenty AI-model sessions tested a shared spending limit. Five in-flight transactions lost their parent authority before commit. The local evidence showed what actually happened.

Read #008 →
When Agents Run Amok series illustration
WHEN AGENTS RUN AMOK #007
NEW · OCTOBER 8, 2026

Kill Switch vs. Stop Sign

Kill switches matter. But sometimes the agent is fine and the action is not. #007 looks at the difference between shutting an agent down and stopping the wrong action before execution.

Read #007 →
When Agents Run Amok #007: Kill Switch vs. Stop Sign
WHEN AGENTS RUN AMOK#006
SEPTEMBER 23, 2026

The Work Was Authorized. Then It Wasn't.

The worker, credential, and queued operation can remain unchanged while the originating authority disappears. #006 asks why execution-time revalidation matters.

Read #006 →
When Agents Run Amok series graphic
WHEN AGENTS RUN AMOK#005
SEPTEMBER 16, 2026

An Agent Should Never Be Able to Approve Itself

An approval artifact is not proof of authority merely because it exists. #005 asks whether the actor requesting power can also be trusted to create the evidence that grants it.

Read #005 →
When Agents Run Amok series graphic
WHEN AGENTS RUN AMOK#004
SEPTEMBER 3, 2026

The Problem With Proving an Agent Was Wrong

Logs can reconstruct what an agent did after the fact. The harder question is whether the agent had legitimate, current, delegated authority before the action occurred.

Read #004 →
When Agents Run Amok series graphic
WHEN AGENTS RUN AMOK#003
AUGUST 31, 2026

The Problem With Telling an Agent No

A denial can stop an action. It can also tell an adaptable agent where the boundary is. #003 follows Amok as it treats DENY as information and keeps working the problem.

Read #003 →
When Agents Run Amok series graphic
WHEN AGENTS RUN AMOK #002
AUGUST 26, 2026

Who actually authorized the agent?

An agent can be authenticated, hold valid credentials, and still take an action nobody actually authorized. The second installment looks at why identity and access stop short of answering the authority question.

Read #002 →
When Agents Run Amok series graphic
WHEN AGENTS RUN AMOK #001
INCIDENT ANALYSIS

The credential remained valid. The authority did not.

A compromised agent can remain authenticated while its actions move beyond the authority originally delegated. This installment examines the control gap.

Read #001 →
PRIMER

Identity, access, and authority are not the same thing.

Identity establishes the actor. Access determines what a credential can reach. Authority asks whether the actor is legitimately empowered to take this specific action, for this purpose, at this time.

Praeva is built around the third question. It is not intended to replace IAM, authentication, authorization policy, or endpoint detection. It is designed to evaluate legitimate delegated authority at the execution boundary.

PUBLIC DOCTRINE

Five principles for delegated agent authority.

  1. Authority originates outside the model. Models do not create their own legitimate power.
  2. Delegation should be explicit. Purpose, scope, actions, constraints, expiry, and lineage should be inspectable.
  3. Authority should be locally verifiable. Consequential systems should not rely on vague upstream assumptions.
  4. Authority must be revocable. Runtime decisions must respond when delegation expires or changes.
  5. Decisions should leave evidence. Organizations should be able to reconstruct the authority chain and decision rationale.