Understanding AI Resumability

From Hydration to Resumability

Learn how the concept of "resumability" from web frameworks like Qwik can revolutionize how AI systems manage cognitive state.

The Problem with Traditional AI

Hydration (Current)

Replay-based approach

Rebuilds context every time

Replays entire conversation history

High token consumption

Re-processes all previous reasoning

Loses intermediate state

No access to reasoning structure

Context window limits

Eventually truncates important history

Process Flow

LoadParseRebuildRe-deriveRe-planReady

Resumability (Future)

Restore-based approach

Restores state directly

Loads serialized cognitive checkpoint

Minimal token usage

Only processes new interactions

Preserves reasoning structure

Intent graphs, decisions, tool state

Enables branching

Fork and explore alternative paths

Process Flow

LoadRestoreReady

The Continuation Record (CR)

The CR is the fundamental artifact that enables cross-agent resumability—capturing not just state, but the full context needed for any thinking agent to continue work seamlessly.

CR Structure

ContinuationRecord {
  // Identity & Audit Trail
  id, version, contentHash
  parentCRHash, merkleRoot

  // Task Definition
  task { goal, successCriteria, status }

  // Execution Summary
  executionSummary {
    stepsCompleted, tokensConsumed
  }

  // Decisions & Learnings
  decisions: CRDecision[]
  learnings: CRLearning[]

  // Human Interaction
  questionsForUser: CRQuestion[]
  blockers: CRBlocker[]
  resumePoints: CRResumePoint[]

  // Tool Effects (idempotency)
  toolEffects: CRToolEffect[]
}

Key Properties

Hash-linked chain

Every CR links to its parent via content hash, creating an immutable audit trail

Multi-agent ready

Any agent (LLM, human, deterministic) can resume from a CR

Human-in-the-loop

Questions, approvals, and overrides built into the structure

Tool idempotency

Track what tools ran, their effects, and whether they can be replayed

CR Components

The building blocks that make up a Continuation Record

Resume Points

Explicit continuation gates with preconditions, authority requirements, and approval tracking. Define when and how work can continue.

CRResumePoint

Decisions

Tracked choices with status (tentative/accepted/rejected/contested), rationale, and the ability to be overridden by humans.

CRDecision

Questions

Human intervention requests: questions for clarification, approvals needed, override requests, or scope changes.

CRQuestion

Learnings

Insights, constraints, patterns, and warnings captured during execution. Can come from observation, failure, success, or feedback.

CRLearning

Tool Effects

Record of tool invocations with inputs, outputs, idempotency keys, and whether they're replayable. Prevents duplicate side effects.

CRToolEffect

Blockers

Obstacles preventing progress: missing info, permissions needed, dependencies, errors, or human-required decisions.

CRBlocker

Two-Layer Architecture

How internal reasoning state relates to the external handoff artifact

Internal State

How AI thinks

CognitiveNode

Individual reasoning steps

EpistemicState

Facts, beliefs, questions

IntentGraph

Goals hierarchy

compiles to

External Artifact

What gets shared

ContinuationRecord

Complete handoff package

Contains: decisions, questions, resume points, tool effects, learnings, blockers

CognitiveNode (Internal)

• Fine-grained reasoning steps
• Ephemeral processing state
• Graph of thoughts with dependencies
• Used during active execution

ContinuationRecord (External)

• Structured handoff artifact
• Persistent, serializable, portable
• Hash-linked audit trail
• Agent-agnostic: any AI or human can resume

"The future isn't better prompts. It's persistent, resumable cognition."
— The core insight of resumable AI systems

Benefits of the CR System

90%+ Token Savings

Restore from a CR instead of replaying conversation history. A 45,000 token context rebuild becomes ~1,200 tokens to load.

Near-Instant Resume

The CR contains everything needed: decisions made, questions answered, tool effects recorded. No re-derivation needed.

Multi-Agent Handoff

Pass a CR between GPT-4, Claude, human reviewers, or deterministic systems. The CR is agent-agnostic.

Hash-Linked Audit Trail

Each CR links to its parent via contentHash. Fork, branch, and merge reasoning paths with full provenance.

Human-in-the-Loop Built In

CRQuestion for clarification, CRResumePoint for approvals, CRDecision overrides—all first-class citizens.

Idempotent Tool Execution

CRToolEffect tracks what ran, with idempotency keys. Safe to resume without re-executing side effects.

Dive Deeper

Explore the full CR type system, real examples, and hands-on code

Try It Yourself

Explore the interactive demo to see resumability in action

Launch Interactive Demo Open Playground