{
  "_meta": {
    "publication": "Agent Techniques Weekly",
    "schemaVersion": "2026.05.02",
    "generatedAt": "2026-09-19T18:20:54.892Z",
    "canonicalUrl": "https://brianletort.ai/industry/agents/2026-W38",
    "markdownUrl": "https://brianletort.ai/industry/agents/2026-W38/llm.md",
    "sourceFile": "src/data/industry/agents/2026-W38.ts"
  },
  "issue": {
    "slug": "2026-W38",
    "isoYear": 2026,
    "isoWeek": 38,
    "issueNumber": 22,
    "publishedAt": "2026-09-19",
    "cadence": "weekly",
    "periodLabel": "Week 38 of 2026",
    "bigRead": {
      "headline": "A live agent can now talk while it works, which makes cancellation and state reconciliation the new control",
      "body": "Google's Gemini 3.8 Live release turns a voice turn into concurrent work: the agent can continue the conversation while tools run, and the Extended Thinking variant can reason and narrate progress at the same time. The transferable technique is not voice. It is separating a responsive foreground loop from a durable background task, then giving both a shared cancellation token, an authoritative task state, and a reconciliation step before any result becomes final.\n\nWithout those controls, conversational continuity hides execution risk. A user can interrupt, change the instruction, or revoke permission while a tool call is still running. A narrated plan can become stale before the background task returns. The verifier therefore has to compare the final tool result with the latest user intent and permission state, not only check whether the tool succeeded.\n\nThe week's strongest proof point came from Microsoft rather than a launch benchmark. Microsoft reports more than 111 agents in cloud supply-chain workflows, with average monthly planning cycle time declining from about ten business days to less than 2.5 across five measured cycles, and named the team size, period, and workflow. It remains internal analysis, but it is materially stronger than an unbaselined productivity multiple. The operating lesson is to measure a bounded workflow across repeated cycles and keep the human validation step visible."
    },
    "technique": {
      "name": "Foreground conversation, background execution, shared cancellation",
      "mode": "automate",
      "summary": "Keep the user-facing conversation responsive while tools execute asynchronously, but bind both planes to one task identifier, one current-intent record, one permission snapshot, and one cancellation path.",
      "whyItMatters": "Parallel conversation reduces perceived latency and makes long-running work usable, but it also allows the user's intent and the executing task to diverge. A shared cancellation and reconciliation protocol prevents the agent from completing an obsolete or no-longer-authorized action simply because the tool call was already in flight.",
      "anatomy": [
        {
          "label": "Trigger",
          "detail": "Create a durable task identifier when the user authorizes background work. Record the exact requested outcome, the tools permitted, and the approval level before acknowledging that execution has started."
        },
        {
          "label": "Foreground loop",
          "detail": "Let the conversational model acknowledge, clarify, and narrate without pretending the background result exists. Every progress statement must be derived from task state rather than generated from the original plan."
        },
        {
          "label": "Background executor",
          "detail": "Run tools against a versioned intent and permission snapshot. Require idempotency keys for external writes so retries or resumed sessions cannot duplicate actions."
        },
        {
          "label": "Reconciliation verifier",
          "detail": "Before presenting or committing a result, compare it with the latest user instruction, current permissions, and task version. A technically successful tool call fails verification if the user changed or cancelled the goal."
        },
        {
          "label": "Cancellation and escalation",
          "detail": "Propagate user interruption to queued and running work, stop safely where possible, and surface any irreversible side effect that completed before cancellation. Route ambiguous or high-impact divergence to a human."
        }
      ],
      "examples": [
        "Gemini 3.8 Live maintaining dialogue while tools and APIs execute in the background",
        "Gemini 3.8 Live Extended Thinking narrating multistep progress while deeper reasoning continues",
        "A support agent that keeps speaking while a refund check runs, but revalidates the requested amount before submission"
      ],
      "source": "Google",
      "sourceUrl": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/"
    },
    "agentCapabilities": [
      {
        "vendor": "Google",
        "product": "Gemini 3.8 Live",
        "mode": "cowork",
        "date": "2026-09-15",
        "capability": "Native speech-to-speech interaction with near-real-time visual grounding, automatic movement across 97 languages, and background tool or API execution while dialogue continues.",
        "meaning": "Operators should separate conversational responsiveness from task completion in telemetry and user experience. A smooth foreground must never be accepted as evidence that the background action succeeded, remained authorized, or still matches current intent.",
        "source": "Google",
        "sourceUrl": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/"
      },
      {
        "vendor": "GitHub",
        "product": "Copilot automatic model selection tiers",
        "mode": "build",
        "date": "2026-09-14",
        "capability": "Three routing policies let a user optimize automatic model selection for efficiency, balance, or intelligence while the service chooses a model per prompt from the same available set.",
        "meaning": "Engineering leaders should log the selected model, latency, cost, and outcome for every routed task. The policy name expresses intent, but only the route-level record can explain why spend or quality changed between otherwise similar sessions.",
        "source": "GitHub Changelog",
        "sourceUrl": "https://github.blog/changelog/2026-09-14-configure-cost-and-quality-in-copilot-auto-model-selection/"
      },
      {
        "vendor": "GitHub",
        "product": "Agent Host local Dev Containers and pull-request creation",
        "mode": "build",
        "date": "2026-09-18",
        "capability": "Agents can run with project tools and dependencies inside supported local Dev Containers, create pull requests directly from sessions, and optionally clean up inactive sessions after their pull requests merge.",
        "meaning": "A reproducible container narrows environment drift, but local execution inherits the developer machine's credentials and network reach. Pair the container definition with managed permissions, an explicit outbound-domain policy, and retained session evidence before enabling direct pull-request creation.",
        "source": "GitHub Changelog",
        "sourceUrl": "https://github.blog/changelog/2026-09-18-github-copilot-weekly-releases-september-14/"
      },
      {
        "vendor": "Salesforce",
        "product": "AIforce Headless Toolkit",
        "mode": "cowork",
        "date": "2026-09-15",
        "capability": "CRM data, workflows, business logic, semantics, permissions, security, and governance are exposed to Claude, Slack, Lightning, and partner-built agent interfaces.",
        "meaning": "The agent interface can change without moving the system of record, but permission enforcement and action logging must remain consistent across surfaces. Test the same restricted action through each interface and reconcile the resulting audit trail.",
        "source": "Salesforce",
        "sourceUrl": "https://www.salesforce.com/news/stories/aiforce-announcement/"
      }
    ],
    "skillsAndConnectors": [
      {
        "ecosystem": "Salesforce",
        "name": "AIforce Headless Toolkit and AgentExchange",
        "type": "connector",
        "date": "2026-09-15",
        "signal": "A headless layer exposes MCPs, APIs, plugins, skills, workflows, and actions so third-party AI interfaces can use Salesforce context and Agentforce agents.",
        "why": "This makes the connector boundary more durable than the interface. Platform teams should inventory which permissions, semantic definitions, and audit records survive across Claude, Slack, Lightning, and partner surfaces before approving broad activation.",
        "source": "Salesforce",
        "sourceUrl": "https://www.salesforce.com/news/stories/aiforce-announcement/"
      },
      {
        "ecosystem": "Gemini Live API",
        "name": "Real-time media integration partners",
        "type": "connector",
        "date": "2026-09-15",
        "signal": "Google lists Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, and Vision Agents as integration partners handling media-streaming infrastructure for Gemini Live applications.",
        "why": "Media transport becomes a separate operational dependency from the model and tools. Buyers should record which partner terminates audio, where media is retained, how interruptions propagate, and whether end-to-end latency remains observable.",
        "source": "Google",
        "sourceUrl": "https://blog.google/innovation-and-ai/technology/developers-tools/build-real-time-voice-applications-gemini-audio/"
      },
      {
        "ecosystem": "GitHub",
        "name": "Sentry integration in Copilot",
        "type": "connector",
        "date": "2026-09-18",
        "signal": "The weekly Copilot release adds Sentry integration alongside code-review and agent-session updates, bringing production error context closer to the build agent.",
        "why": "Production telemetry can improve diagnosis, but it also expands the sensitive context available to an agent. Scope projects and fields explicitly, redact secrets, and verify that write authority does not arrive with read access.",
        "source": "GitHub Changelog",
        "sourceUrl": "https://github.blog/changelog/2026-09-18-github-copilot-weekly-releases-september-14/"
      }
    ],
    "proofOfValue": [
      {
        "actor": "Microsoft",
        "workflow": "Cloud supply-chain monthly planning and demand-plan investigations",
        "evidence": "vendor_claim",
        "claim": "Across five monthly planning cycles from April through August 2026, average cycle time declined from about ten business days to less than 2.5 after more than 111 agents were deployed.",
        "read": "This is stronger than a generic productivity percentage because Microsoft names the workflow, period, team size, cycle count, baseline, and resulting duration, and states that explanations remained human-validated. It remains internal analysis without a control group, so use the shape of the measurement rather than importing the magnitude into a business case.",
        "source": "Microsoft",
        "sourceUrl": "https://blogs.microsoft.com/blog/2026/09/17/what-weve-learned-from-microsofts-own-ai-transformation/"
      },
      {
        "actor": "Salesforce",
        "workflow": "CRM actions such as updating opportunities, routing cases, and scheduling follow-ups",
        "evidence": "vendor_claim",
        "claim": "Salesforce reports that Koa matches or exceeds leading models on its CRM benchmark with three times fewer errors.",
        "read": "The named tasks make the claim testable, but the benchmark, absolute error rates, comparison roster, and evaluation method are not published. Reproduce the task set in the pilot and require absolute failure counts before using the multiple.",
        "source": "Salesforce",
        "sourceUrl": "https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/"
      }
    ],
    "enterpriseReadiness": [
      {
        "dimension": "permissioning",
        "note": "Snapshot permission at task start and revalidate it before any external write. A live conversation can outlast the authorization that launched its background tool, especially when users switch accounts, roles, or intent mid-turn."
      },
      {
        "dimension": "verification",
        "note": "Verify against the latest task version, not the original prompt. A successful tool result is invalid if the user cancelled, changed the requested outcome, or supplied newer evidence while the background work was running."
      },
      {
        "dimension": "auditability",
        "note": "Use one task identifier across foreground dialogue, model routing, tool calls, approvals, retries, and final disposition. Concurrent execution is not auditable when each plane emits an unrelated session log."
      },
      {
        "dimension": "cost",
        "note": "Meter the entire workflow: audio input, audio output, reasoning, tools, retries, media transport, and idle open-session time. A low voice rate can coexist with an expensive background execution path."
      },
      {
        "dimension": "reliability",
        "note": "Require idempotency keys and explicit cancellation semantics for every side-effecting tool. Retrying or resuming a background task must not duplicate a payment, ticket, message, or production change."
      }
    ],
    "scorecard": {
      "asOf": "2026-09-19",
      "rows": [
        {
          "mode": "chat",
          "leadingPattern": "Live multimodal conversation with explicit per-minute meters and background tools",
          "representativeTools": [
            "Gemini 3.8 Live",
            "GPT-Live-1"
          ],
          "controlGap": "Foreground fluency can conceal stale, failed, or cancelled background work unless task state is surfaced separately."
        },
        {
          "mode": "cowork",
          "leadingPattern": "System-of-record context and actions exposed through interchangeable AI interfaces",
          "representativeTools": [
            "Salesforce AIforce",
            "Agentforce Coworker",
            "Claudeforce"
          ],
          "controlGap": "Cross-surface permission equivalence and audit-log reconciliation are asserted but need tenant-level testing."
        },
        {
          "mode": "build",
          "leadingPattern": "Policy-based model routing inside reproducible project environments",
          "representativeTools": [
            "GitHub Copilot auto tiers",
            "GitHub Agent Host Dev Containers",
            "Cursor Projects"
          ],
          "controlGap": "Routing intent is visible, but selected-model cost and outcome need route-level export for governance."
        },
        {
          "mode": "automate",
          "leadingPattern": "Durable background work with a responsive foreground and shared cancellation",
          "representativeTools": [
            "Gemini Live background tools",
            "Temporal durable execution",
            "Microsoft supply-chain agents"
          ],
          "controlGap": "Most launch material does not specify idempotency, cancellation propagation, or reconciliation against changed intent."
        }
      ]
    },
    "tryThis": {
      "title": "Interrupt a background task and prove nothing stale can commit",
      "steps": [
        "Choose a harmless workflow with a delayed tool call, such as drafting but not sending a status message, and assign one durable task identifier to the conversation and tool execution.",
        "Start the task, then interrupt with a materially changed instruction before the tool returns. Record the new task version and issue cancellation to the old version.",
        "Let the original tool call return and confirm the verifier rejects its result because it no longer matches current intent, even though the tool itself succeeded.",
        "Repeat with a permission revocation and with a network retry. Confirm the idempotency key prevents duplicate side effects and the audit record links both attempts."
      ],
      "expectedOutcome": "You will have evidence that conversational continuity cannot cause an obsolete task to commit, plus a concrete gap list for cancellation propagation, permission revalidation, idempotency, and cross-plane audit correlation."
    },
    "watchlist": [
      {
        "window": "Sep 21-30",
        "title": "Gemini Live cancellation and session semantics",
        "why": "Developer documentation needs to state how background tool calls behave when the user interrupts, disconnects, or changes intent."
      },
      {
        "window": "Oct 2026",
        "title": "Koa pilot error evidence",
        "why": "Absolute task-level failure counts would turn a vendor multiple into an actionable verification target."
      },
      {
        "window": "Next GitHub release",
        "title": "Route-level cost and selected-model telemetry",
        "why": "Policy tiers become governable only when organizations can export which model ran, what it cost, and whether the task succeeded."
      },
      {
        "window": "Q4 2026",
        "title": "Microsoft supply-chain measurement update",
        "why": "Additional planning cycles would show whether the reported duration reduction persists beyond five observations."
      }
    ],
    "changelog": [
      "Lead technique changed from independent verification in W37 to shared cancellation and reconciliation for concurrent conversational execution.",
      "Proof quality remains explicit: Microsoft's bounded internal analysis and Salesforce's proprietary benchmark are both labeled vendor claims.",
      "All product and pricing claims are sourced to public material dated September 14-19, 2026."
    ]
  }
}