Release notes

Changelog

New resources and site improvements, recorded when they ship.

  1. Content refresh

    Adopt managed agent sessions without surrendering control

    A new AI-tooling lesson and operational workflow map OpenAI's public-beta Agents API to application-owned authorization, durable side-effect evidence, retention decisions, recovery tests, and release gates.

    AI toolingManaged agent API adoptionUnderstand which runtime responsibilities the managed harness takes on and which production controls remain with your application.WorkflowManaged Agent Session ReleaseInventory the workload, version the session contract, persist evidence, test recovery, and canary the managed execution path.

    6 glossary additions

    Managed Agent HarnessAgent SessionAgent Execution EnvironmentSession EventAgent VaultZero Data Retention

    Researched September 15, 2026 from OpenAI's September 10 Agents API public-beta announcement and official Agents API overview, environment, session, multi-agent, and data-control documentation. Provider-managed capabilities are separated from SLOPSTACK's application-control and release recommendations. Sources: https://openai.com/index/introducing-the-agents-api/ and https://developers.openai.com/api/docs/guides/agents-api/overview

  2. Content refresh

    Run coding agents in parallel without integration chaos

    A new Ryan VerWey guide and reusable planning skill turn large software tasks into dependency-aware work packets with isolated ownership, bounded concurrency, inspectable evidence, and ordered integration.

    GuideParallel Coding Agents Without Integration ChaosDecide what can run concurrently, prevent overlapping changes, manage fan-in, and verify the combined result before release.SkillParallel Agent Work PlannerProduce a dependency graph, exclusive work packets, a concurrency budget, and a final integration plan without dispatching agents.

    6 glossary additions

    Agent Fan-OutFan-In IntegrationTask IndependenceWorkspace OwnershipConcurrency BudgetIntegration Owner

    Researched September 10, 2026 from current GitHub Copilot app and SDK documentation for isolated sessions, sub-agent orchestration, fleet dispatch, and independent critique, plus official OpenAI model guidance for multi-agent delegation. Product capabilities are separated from SLOPSTACK's planning and integration recommendations.

  3. Content refresh

    Move off retiring AI models without surprises

    A new workflow and Ryan VerWey prompt turn model deprecation notices into owned inventories, replacement behavior contracts, reversible rollouts, and evidence of zero retired-model traffic.

    WorkflowAI Model Retirement MigrationFind direct and hidden model dependencies, validate the replacement, stage the cutover, and remove the retired path before shutdown.PromptAI Model Retirement Migration PlannerConvert a provider notice and workload evidence into a sourced migration plan with owners, gates, rollback, and stop conditions.

    6 glossary additions

    Model LifecycleModel DeprecationModel RetirementMigration WindowModel Behavior ContractModel Fallback Policy

    Researched September 8, 2026 after GitHub's September 3 Copilot model retirement notice. The workflow cross-checks GitHub's supported-model history with Google Gemini and Anthropic lifecycle documentation; current primary-source URLs are included in the planner prompt.

  4. Model update

    Claude earns its place at the new frontier

    Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5 join the rankings with benchmark-grounded grades, while the full text and code catalog moves to a stricter, auditable Frontier Fit calculation.

    ModelClaude Fable 5.1Anthropic's highest-capability generally available model for long-running coding, reasoning, and research.ModelClaude Opus 5The recommended daily flagship for complex agentic coding and enterprise work at half Fable's token price.ModelClaude Mythos 5.1The invite-only Fable sibling for vetted cybersecurity and life-sciences organizations.

    Grades use Anthropic's same-harness evaluation table and the independent Terminal-Bench 4.0 leaderboard. Astra and Fable 5.1 are treated as statistically tied on that coding benchmark; Opus 5 remains close, while Sol's previous near-perfect capability grade was reduced. Frontier Fit no longer awards keyword-derived bonuses, and limited access affects overall fit without changing capability bars.

  5. Model update

    GPT-6 Astra resets the frontier baseline

    GPT-6 Astra joins the model rankings as the new capability leader, and every text and code model is recalibrated against the September 2026 frontier.

    ModelGPT-6 AstraA premium model for the hardest coding, reasoning, research, computer-use, and long-horizon agent workflows.

    The assessment uses OpenAI's current model documentation for capabilities, context, output limits, modalities, and token pricing. Capability scores establish a new frontier ceiling; speed and cost remain independently graded, and prior models retain their release-era evidence before the existing generation-age calibration is applied.

  6. Content refresh

    Know what your coding agent can access

    A new AI-tooling lesson and reusable audit skill separate content filtering, MCP server admission, and tool permissions, with safe tests for client-specific policy coverage.

    AI toolingAgent data-access boundariesMap every context path to its enforcing control and verify coverage without exposing protected data.SkillAgent Data Access AuditA read-only audit contract with a coverage matrix, synthetic tests, and explicit unknowns.

    6 glossary additions

    Content ExclusionMCP Server AllowlistPolicy-Surface CoverageIndirect ContextTool AvailabilityFail-Closed Policy

    Researched September 3, 2026 from GitHub's September 2 content-exclusion release and official content-exclusion, enterprise managed settings, and CLI permissions documentation. Source links are included in the audit skill; the testing procedure is SLOPSTACK guidance.

  7. Content refresh

    Risk-calibrated review for agent-authored changes

    A new guide, workflow, and prompt help teams match review depth to change risk, bound automated-review context, adjudicate findings with evidence, and recheck the final commit before merge.

    GuideRisk-Calibrated AI Code ReviewReview agent-authored changes with explicit risk contracts, layered analysis, evidence-backed findings, and independent merge gates.WorkflowRisk-Calibrated Agent ReviewA repeatable path from change-risk classification through final-head re-review and owner approval.PromptAgent Change Review AdjudicatorTurn an agent-authored diff and its review threads into evidence-backed findings, decisions, and merge gates.

    7 glossary additions

    Agentic Code ReviewReview Effort LevelReview AdjudicationAutomatic Re-reviewResolution ReasonReview Context BoundaryBot-authored Pull Request

    The release uses current GitHub Copilot code review documentation and August 2026 announcements for effort levels, bot-authored and large-PR coverage, resolution reasons, automatic re-review, skills, MCP context, and second-opinion review.

  8. Content refresh

    Operational visibility for production agents

    A guide, reusable audit skill, and workflow now connect agent traces, privacy controls, telemetry design, and downstream outcomes into one practical observability loop.

    GuideAgent ObservabilityInstrument agent workflows with stable identities, trace context, bounded telemetry, privacy controls, and outcome evidence.SkillAgent Observability AuditReview an agent system for trace completeness, semantic consistency, cardinality risks, redaction, and outcome correlation.WorkflowAgent Observability LoopA privacy-aware telemetry loop that connects agent operations and adoption metrics to reviewed outcomes.

    7 glossary additions

    Agent ObservabilityGenAI Semantic ConventionsAgent SpanTool Call SpanTrace ContextTelemetry CardinalityTrace Redaction

    The guidance is grounded in OpenTelemetry GenAI semantic conventions and context propagation, OpenAI Agents SDK tracing controls, and GitHub Copilot usage-metrics documentation.