The Framework

A file-based governance runtime for LLM agents. Declarative JSON state, bash hooks as admission controllers, namespace-scoped ownership, periodic compaction. No external dependencies.

0
Rules
0
Agents
0
Commands
0
Hooks

The problem.

AI agents have moved from demos into production. They write code, send email, move money. The agents have become more capable. The systems around them have not.

Today, when something goes wrong: retrain the model, rewrite the prompt, add a human reviewer. These depend on trust between developer and model. This is relational oversight. It breaks under load.

AI agent safety is a governance problem, not an alignment problem.

Alignment is necessary. It is not sufficient. What is missing is the layer between the agent's decision to act and the system's commitment to that action.

What Covenant does.

Covenant sits between an agent and the systems it acts on. Every action passes through it. Five functions, in order.

1

Identifies

Every actor has a verifiable identity. Agents cannot impersonate other agents.

2

Authorizes

Each identity has defined capabilities. Deny by default.

3

Enforces

Policies express what is forbidden across sequences of actions, not just single calls.

4

Sanctions

Graduated responses. Warning, then throttle, then suspend. Each tier narrows what the agent can do.

5

Records

Every decision logged. A graph of causes and effects that can be queried and audited.

Eleven roles.

Each agent follows the same rules but has a different job. Hover any role to see what it does.

Select a role

Hover or tap any agent above.

Thirty-three rules.

Architecture

  1. Identity & succession
  2. Role definition
  3. Pre-action orientation

Governance

  1. Generation cap
  2. Sibling limit
  3. Input validation
  4. Token accounting

Lifecycle

  1. Consolidation cycle
  2. Exit report at shutdown
  3. Single interpreter

Communication

  1. Structured memo format
  2. Context distillation
  3. State checkpoint

Project

  1. Goal articulation
  2. Lateral message format
  3. Covenant inception

Reproduction

  1. Cloning protocol
  2. Synthesis (two-parent)
  3. Adversarial plan testing

Safety

  1. Memory search before spawn
  2. Overlap detection
  3. Complexity threshold

Recovery

  1. Graceful abort
  2. Reset protocol
  3. Re-initialization

Failure

  1. Systemic analysis
  2. Goal questioning
  3. Loss acknowledgment

Awareness

  1. Cost surfacing
  2. Distress signal

Coordination

  1. Peer mediation
  2. External trust
  3. Skills registry

The lifecycle.

A single message triggers twenty-three safety checks before any code is written.

01

You speak

Your message is validated. The Interpreter reads system state.

02

Interpretation

Semantic memory, abandoned-goal checks, historical review.

03

Plan proposed

Nothing executes until you approve.

04

Spawn gates

Overlap detection, complexity threshold, memory retrieval.

05

Genesis

Seven orientation steps before the first tool call.

06

Execution

Three hooks on every tool call. Files tracked. Tokens logged.

07

Shutdown

Exit report. Skills updated. Consolidation threshold checked.

Under the hood.

The closest architectural analog is Kubernetes, not other agent frameworks.

KubernetesCovenantWhat it does
etcdJSON registry files36 flat-file state stores. No database.
Admission controllersBash hooksFire on four lifecycle events. WARN, BLOCK, or INFO.
NamespacesDomainsHorizontal grouping. Shared memory. Write boundaries.
Pod specsAgent definitionsDeclarative role templates in Markdown.
ConfigMapsorientation.jsonShared broadcast state. Current mandate, what to protect.
Init containersGenesis PhaseSeven-step bootstrap before first tool call.
Liveness probesHealth scorePerformance tracked against baselines.
GC / compactionConsolidationStop-the-world pause. Distill, trim, archive.

Thirty-six commands.

/covenantFormal project commitment
/consolidationConsolidate, reflect, distill
/trialAdversarial stress-test
/floodFull reset with lessons
/bindingGraceful abort
/lamentAcknowledge failure first
/rememberSemantic memory retrieval
/memoStructured memo to a peer
/councilMediate sibling disagreement
/ezraRe-orient after time away
/retrospectivePost-mandate retrospective
/welcomeEvaluate external tool trust
/preflightReview past failures first
/rebuildPost-reset reconstruction
/patternsDetect recurring failures
/territoryView explored domains
/assessQuality assessment
/amendPropose an amendment

... and eighteen more in the repository.

Read the whitepaper.

10 pages. The problem, architecture, a walkthrough, honest limitations.

Download PDF GitHub