Skip to content
Press / to search

ContextOS Research Synthesis

A traceable evidence-to-spec crosswalk showing how all 108 ContextOS research essays evolve the five-plane runtime without creating competing contracts.

Research SynthesisLast reviewed: Edit on GitHub
At a glance

This page records how the complete ContextOS blog corpus was absorbed into the documentation. The snapshot covers 108 essays published through 2026-08-24. It is both a research index and a change-control boundary: each essay has a canonical destination, while proposals that are not yet runtime contracts remain visibly non-canonical.

The result is not a sixth plane and not a second specification. It is a traceable promotion path from observations and experiments into the existing five-plane model.

Evidence-to-spec promotion rule

Blog language is promoted at one of three levels:

LevelMeaningDocumentation treatment
Canonical contractImplemented and represented by the typed runtime contract.Use the exact type and field names from src/lib/contextos/types.ts and the schema reference.
Documented protocolA durable operating pattern composed from canonical contracts.Specify the invariants, evidence, gates, and metrics without pretending it is a new runtime type.
Candidate extensionA useful proposal whose schema or semantics are not yet settled.Name it as a candidate, state which existing contracts bound it, and require an ADR plus conformance tests before promotion.

This distinction matters. A “release manifest” is currently a documented aggregate of pinned versions and evidence already carried across RunContext, CompiledContext, tool envelopes, traces, and the DecisionRecord; it is not a new core interface. “Autonomy budget,” “irreversibility budget,” and “reversal spec” remain candidate policy or effect-protocol extensions. The canonical controls today are RunBudget, ActionRisk, policy decisions, approvals, idempotency, postconditions, and compensation receipts.

Integrated construct ledger

The corpus adds the following durable requirements to the existing model:

ConstructStatusCanonical homeRequired outcome
Complete release tupleDocumented protocolHarness EngineeringEvery run can identify code, model route, prompts, packs, policies, skills, tools, evaluators, and environment.
Proof-carrying executionDocumented protocolHarness EngineeringA terminal claim is backed by identity, authority, evidence, effects, postconditions, and recovery receipts.
Portable trajectory evidenceDocumented protocolEvaluation and ObservabilityEvals inspect observable decisions, calls, results, state diffs, and verdicts without requiring hidden chain-of-thought.
Paired Skill LiftDocumented protocolAdapter Mesh and EvaluationA skill ships only when it adds measurable value over a no-skill baseline without unacceptable safety, cost, or latency regression.
Lifecycle securityDocumented protocolHarness EngineeringSecurity gates cover configuration, extension, runtime, persistence, action, and recovery phases.
Authority non-escalationCanonical invariantGovernanceContext, memory, tool metadata, and agent messages may inform a decision but cannot grant themselves authority.
Effective authority graphDocumented control viewIdentity LayerOperators can resolve and revoke the actual identity → delegation → capability → resource path.
Recall-time memory reconstructionDocumented protocolMemory ModelStored material is recompiled under current evidence, policy, authority, freshness, and contradiction state before use.
Independent multi-agent certificationDocumented protocolOrchestrationConsensus never substitutes for external evidence, verifier independence, or one accountable runtime owner.
Risk-weighted human oversightDocumented protocolEvaluation and ObservabilityAudit allocation combines deterministic risk gates with stratified random coverage and reviewer-quality measurement.
Effect and recovery receiptsCanonical compositionAdapter MeshExternal action is normalized, idempotent, postcondition-checked, and paired with rollback or compensation evidence where applicable.
Cost per trusted outcomeCanonical metric aliasMetrics GlossaryTotal inference, tool, retry, evaluation, observability, infrastructure, and review cost rolls up to contextos.budget.cost_per_verified_success.
Execution / evaluation / improvement separationCanonical invariantImprovement LoopThe system being optimized cannot define or waive its own release gate.
Evaluated prompt contractDocumented protocolAgentic Context EngineeringPrompts declare trust, authority, tools, output, failure, and stopping behavior and are released with pinned dependencies.

Corpus crosswalk

Each essay below appears once. “Promoted into” names the primary canonical destination; an essay may also inform adjacent pages.

Harness architecture and release engineering

Research essayPromoted constructPromoted into
Adaptive Agent Harnesses with MemoHarnessGoverned adaptation and held-out promotion gatesImprovement Loop
Agent Harness WhitepaperCross-plane harness as the production unitHarness Engineering
Agents.md Done RightRepository-visible operating contractHarness Engineering
Beyond Prompts: Architecture of TrustControl outside model behaviorFoundations
ContextOS Research PaperGoverned runtime thesis and five planesReference Architecture
Develop an Agent with a Harness End to EndFull execution contractHow It Works
Eight-Property Harness AuditProduction-readiness controls with evidenceHarness Engineering
Five Planes of Agentic Operating SystemsPlane ownership and typed seamsFoundations
Great AI Engineers Build Agent HarnessesHarness, not prompt, as engineering scopeHarness Engineering
Harness Engineering Field Report, Aug 3–13Field evidence for release/eval/security convergenceHarness Engineering
Harness Engineering Roadmap 2026–2027Release tuple, proof packet, lifecycle roadmapHarness Engineering
Harness Engineering State of the FieldExecution/evaluation/improvement separationHarness Engineering
Production Harness RoadmapStaged adoption across coding-agent runtimesHarness Engineering
State of AI Agents 2026Durable runtime requirements beyond demosContextOS 2026
What Production Agent Runtimes Teach ContextOSTwelve production runtime lawsHarness Engineering

Evaluation, replay, security evaluation, and improvement

Research essayPromoted constructPromoted into
Agent Benchmarks Measure the HarnessRelease-tuple attribution and harness-aware baselinesEvaluation and Observability
Agent Evaluation After LeaderboardsProtocol cards, operating profiles, causal rolloutEvaluation and Observability
Agent Hijacking Security EvalsBoundary-focused adversarial evaluationSecurity and Compliance
Agent Skill LiftPaired baseline and trajectory-graded skill releaseEvaluation and Observability
Autotune the HarnessBounded candidate searchImprovement Loop
Building a Compliance ReviewerEvidence-bound specialist evaluatorReviewer Agents
Building a Reliability ReviewerFailure/recovery specialist evaluatorReviewer Agents
Dataset-First Agent EngineeringDatasets before prompt iterationEvaluation and Observability
Failure Playbooks and Typed VerdictsDeterministic verdict-to-recovery mapFailure Playbooks
From Correction to StrategyRuleTyped correction promotionImprovement Loop
Harness Candidates as Model CheckpointsImmutable candidates and comparative promotionImprovement Loop
Harness Instruction-Following EvalsInstruction hierarchy and end-to-end gradingEvaluation and Observability
Harness Lifecycle SecurityPhase-by-severity security gateHarness Engineering
Human Oversight for Agent FleetsRisk gates, random coverage, reviewer non-vacuityEvaluation and Observability
Pack Rollout in Five StagesShadow-to-full staged rolloutEvaluation and Observability
Replay Harness in CodeExecutable replay contractEvaluation and Observability
Replay Is the Real Audit LogReconstructable evidence over prose logsEvaluation and Observability
Replayable EnvironmentsResettable task fixtures for causal evidenceImprovement Loop
Scorecards Over VibesMulti-axis release scorecardsEvaluation and Observability
Trace Review: The Agent DebuggerCausal trace reviewEvaluation and Observability
Wiring the Five EvaluatorsPolicy, utility, latency, safety, economics wiringEvaluation and Observability

Context compilation, prompts, knowledge, and memory

Research essayPromoted constructPromoted into
Agent Memory Is a Systems WorkloadLifecycle, storage, recall, and repair as one systemMemory Model
AI Agent Memory Is BrokenPromotion over transcript accumulationMemory Model
Antahkarana StackLocal cognitive substrate mappingCognitive Core
Build the Context Pack CompilerDeterministic staged compilationCognitive Core
Context Engineering in ProductionContext as compiled artifactAgentic Context Engineering
Context Graphs in ContextOSEvidence-bound relationship retrievalKnowledge Graph
Context Packs in PracticePack authoring and versioningContext Pack
LLM Latency EngineeringTTFT, prefix caching, routing, and budget tradeoffsAI Gateway and LLM Router
Memory Reconstruction with MemHarnessRecall-time reconstruction under current controlsMemory Model
Persistent Memory for Coding AgentsPortable governed memoryMemory Model
Persistent Memory Poisoning LifecycleCapture, promotion, recall, action, repair boundariesMemory Model
Production Prompt Engineering as Evaluated ContractSeven-part prompt contract and release manifestAgentic Context Engineering
Promotion-Aware Memory in CodeTyped tier transitions and consentMemory Fabric
Prompt Engineering and Latency Field ReportPrompt/cache topology as runtime designAgentic Context Engineering
Reverse PromptingDerive an evaluable contract from the desired receiptAgentic Context Engineering

Decisions, orchestration, and multi-agent systems

Research essayPromoted constructPromoted into
AI Agent Accountability MatrixOne owner per failed decision and control surfaceOrchestration
Graph Engineering for Agent WorkflowsTyped graph states, guards, and recoverable transitionsOrchestration
Multi-Agent Consensus Is Not CorrectnessIndependent evidence, minority preservation, external certificationOrchestration
Passive Awareness with AgentRadioBounded asynchronous evidence channelOrchestration
PM Multi-Agent Control TowerFleet ownership and visible coordination stateOrchestration
Stop Building Multi-Agent Systems Until ReplayReplay before fleet scaleOrchestration
The Critic: Verify, Score, ConsolidateBounded critic contractOrchestration

Action, identity, governance, and security

Research essayPromoted constructPromoted into
AI Agent Access GraphEffective authority graph, drift, and graph revocationIdentity Layer
AI Agent Security Threat ModelThreats mapped to deterministic boundariesSecurity and Compliance
Agent Identity Is the Trust BoundaryWorkload identity and principal chainIdentity Layer
Approval Gates in CodeFrozen evidence and resumable typed interruptGovernance
Approval Mode TiersCompatibility taxonomy bounded by ActionRiskGovernance
Autonomy BudgetComposite policy envelope, not a scalar permissionGovernance
Build the Tool GatewaySingle mediated path to external effectAdapter Mesh
Identity Layer for Agent RunsSigned identity claims and credential exchangeIdentity Layer
MCP Adapters in ProductionProtocol normalization behind the gatewayMCP Standard
Prompt Injection Is a Boundary ProblemInformation cannot grant authorityGovernance
Reversibility Is the Missing Safety PrimitiveEffect postconditions, compensation, and candidate reversal contractGovernance
Secure MCP Tool Supply ChainAdmission, runtime validation, release, kill switchAdapter Mesh
The Glass RuntimeHuman agency through inspectable control and receiptsGovernance

Product contracts, operating models, and domain patterns

Research essayPromoted constructPromoted into
Agentic AI Before and After ContextOSGoverned operating-model contrastOverview
Agentic Incident Command CenterHigh-risk remediation workflowIncident Command Center
Agentic Recommender SystemsEvaluation and policy around recommendation actionsUse Cases
Business Leaders’ Agentic AI Mental ModelRuntime accountability modelOverview
Business Teams: Feedback Loop After LaunchOperator corrections and release evidenceImprovement Loop
Business Teams: Scorecards, Not VibesOutcome scorecards by workflowMetrics Glossary
Business Teams: Trust, Approvals, ReceiptsVisible approval and receipt UXGovernance
Business Teams: Map Real Work Before AgentsWork map before automationIntent-Task Catalog
End-to-End Refund WalkthroughConcrete governed action flowWorkflow Examples
Financial Crime OpsEvidence-rich regulated caseworkFinancial Crime Ops
PM Agent Scorecards and Launch GatesProduct release gatesEvaluation and Observability
PM Operating Agent ProductsPost-launch improvement ownershipImprovement Loop
PM: PRD to Intent CatalogProduct requirements to typed intentsIntent-Task Catalog
PM Trust and Approval DesignApproval UX as a product contractGovernance
Product Managers Build Complex Agentic SystemsProduct control surfaces across planesFoundations
Software Delivery SquadHuman-supervised software workflowSoftware Delivery Squad
Tokenomics: Cost per Trusted OutcomeFull-cost verified-success denominatorMetrics Glossary

External platform and research maps

These essays are evidence about the surrounding field. They update interoperability, model-routing, security, and adoption assumptions; they do not define ContextOS contracts by vendor analogy.

Research essayPromoted constructPromoted into
AI Gateway and LLM Router in ProductionPolicy-bound provider routingAI Gateway and LLM Router
Anthropic 2026 Full-Stack MapExternal platform capability mapContextOS 2026
Anthropic Business Strategy 2026Enterprise adoption and governance evidenceInvest Early
Anthropic Claude Platform and Managed AgentsManaged runtime interoperabilityAdapter Mesh
Anthropic Latest Work, Aug 4–7Safety/runtime research evidenceSecurity and Compliance
Anthropic Research: Interpretability, Science, EconomyResearch signal, not runtime authorityResearch Synthesis
Anthropic Safety, Alignment, and ContainmentLayered containment evidenceSecurity and Compliance
ChatGPT Work, Codex, and OpenAI PresenceProduct-surface/runtime separationContextOS 2026
Claude 5 Model Family Field GuideModel profiles remain router inputsAI Gateway and LLM Router
Claude Code, Cowork, and the Agent Product StackProduct-to-runtime capability mapContextOS 2026
DeepSeek Harness Architecture Security ReviewVendor harness evidence and threat reviewSecurity and Compliance
GPT-5.6 Model Family Field GuideModel-specific behavior isolated in profiles/evalsAI Gateway and LLM Router
OpenAI 2026 Full-Stack Research MapExternal stack evidenceContextOS 2026
OpenAI Latest Work, Jul 29–Aug 7Current research signalsResearch Synthesis
OpenAI Multimodal StackMultimodal evidence and effect boundariesContextOS 2026
OpenAI Responses API and Agent PlatformProvider API behind stable gateway contractsAI Gateway and LLM Router
OpenAI Safety InfrastructureDefense-in-depth evidenceSecurity and Compliance
OpenAI Science, Research, and EvalsExternal evaluation methodsEvaluation and Observability
OpenWorker ReviewMissing trust-runtime controls as negative evidenceHarness Engineering
SecondBrain Local-First Agent OSReference implementation mappingSecondBrain

Conflicts and non-canonical proposals

The synthesis found useful terms that must not be blended into the contract silently:

  • ApprovalMode versus ActionRisk: ApprovalMode remains the v1 compatibility projection. New reasoning uses the independent ActionRisk dimensions. Neither becomes a single numeric “risk score.”
  • Release manifest: required as an immutable aggregate for attribution and replay, but not yet a core interface. Implementations may materialize one if it is derived from and consistent with the canonical pins.
  • Trajectory / protocol card: required evaluation evidence shape, but not hidden chain-of-thought and not yet a core TypeScript type. Store observable envelopes, results, diffs, metrics, and grader provenance.
  • Autonomy and irreversibility budgets: useful policy composition ideas. They do not supersede RunBudget, ActionRisk, approval gates, or deterministic policy checks.
  • Reversal specification: an effect protocol may declare verification, rollback, or compensation. A new ReversalSpec schema requires a separate ADR and compatibility tests.
  • Vendor constructs: provider APIs, model families, and managed-agent products are evidence inputs. They remain behind ContextOS model profiles, adapters, and policy boundaries.

Maintenance contract

When a new research essay is published:

  1. Add it once to this crosswalk and name its primary canonical destination.
  2. Classify the promoted idea as a canonical contract, documented protocol, or candidate extension.
  3. Update the smallest coupled documentation surface; do not create a parallel plane or synonym for an existing type.
  4. If the proposal changes a runtime contract, require an ADR, types/schema updates, examples, migration notes, and conformance tests together.
  5. Record the evaluation, security, replay, and rollback evidence needed to promote it.

The documentation evolves when evidence improves the operating model—not merely because a new term is memorable.