AgentCore lifecycle hooks, LangSmith Tool Gateway, AlphaProof Code verification

Nov 19, 2026•Items: 10

Systems & planning

  • Amazon Bedrock AgentCore adds Strands defaults and…

    TL;DR: Amazon Bedrock AgentCore harness adds Strands harness defaults, context management, built-in tools, and lifecycle hooks for deterministic checks in the agent loop.

    • Automatic prompt caching for supported models using Converse on Bedrock covers system prompts, tool definitions, and conversation history.
    • Context management summarizes older turns while retaining recent context; the source reports about 34% lower token cost than other leading harnesses.
    • Built-ins include plans restored across invocations in a session, web retrieval, granular filesystem access, and delegation; subagents share limits and tool hooks.
    • Four lifecycle checkpoints support synchronous Lambda allow/deny decisions; denied pre-tool calls are skipped, while SNS and EventBridge only notify.

    Why it matters: The synchronous allow/deny checkpoints let you block tool execution for both parent agents and delegated subagents before side effects occur. In contrast, post-invocation checks cannot undo already-streamed model output, so consequential-action policies need to run at or before the pre-tool checkpoint.

    Signals: For a fixed security or capability boundary, explicitly list the required tools in allowedTools; the recommended built-in tool set can gain capabilities over time.

    Source: https://aws.amazon.com/blogs/machine-learning/amazon-bedrock-agentcore-harness-adds-strands-harness-defaults-built-in-tools-and-lifecycle-hooks/

  • Skyflow and AgentCore Gateway de-identify agent tool…

    TL;DR: Skyflow’s Amazon Bedrock AgentCore Gateway interceptor provides configurable de-identification for tool, model, and document data.

    • REQUEST/RESPONSE interceptors use Detect to tokenize or restore entities; Parameter Store policies set direction, entity types, and token type per tool.
    • Shield the tool protects tool arguments and results; Shield the model de-identifies model inputs and re-identifies answers before the agent reads them; Shield the document swaps signed URLs for clean-copy links.
    • If de-identification fails, the call is withheld; file fetching is allowlisted, and CloudWatch logs decisions without credentials or original values.
    • String de-identification is generally available; signed-URL document de-identification is in preview. PDFs/images use one-way tokens; text/Office files use reversible tokens.

    Why it matters: Attaching the interceptor to AgentCore Gateway applies data-protection handling to calls passing through it across agents and tools, without changing agent or tool code.

    Signals: Validate document-de-identification latency against your workloads and Region during a pilot before widening rollout.

    Source: https://aws.amazon.com/blogs/machine-learning/protect-agent-tool-calls-with-skyflow-and-amazon-bedrock-agentcore-gateway/

  • LangSmith Tool Gateway governs MCP tool calls through…

    TL;DR: LangSmith Tool Gateway’s Public Beta routes MCP calls through one URL with identity, access policies, approvals, and tracing.

    • Search mode queries large kits and loads only tools relevant to the current task, rather than loading hundreds or thousands of tools at once.
    • Calls can use the user's credentials or agent credentials; workspace profiles let admins configure shared credentials for background agents.
    • Deny rules can target users or service keys, with exceptions, at org or workspace scope; approval rules require confirmation for sensitive calls.
    • Traces include the tool, server, user, outcome, and duration; tool inputs and outputs are not stored. Enterprise audit logs cover the organization.

    Why it matters: Managing identity and access across growing numbers of users and tools becomes harder at scale. Tool Gateway centralizes tool access and identity resolution, avoiding repeated authentication plumbing for each agent.

    Signals: The post says to contact the LangSmith team for access to deny policies and the audit log.

    Source: https://www.langchain.com/blog/langsmith-tool-gateway

  • HINT-WAM raises LIBERO-90 success to 85.18% with…

    TL;DR: HINT-WAM distills a teacher with hindsight on the student’s own sampling trajectories, reaching 85.18% success on LIBERO-90.

    • The energy score matches teacher action distributions from student samples, not velocity fields; training needs no rewards or simulator interaction.
    • With replay, HINT-WAM reaches 95.10% average on original LIBERO suites and 81.56% on LIBERO-90; without replay, the original-suite average is 44.05%.

    Why it matters: HINT-WAM trains on the student’s encountered states while using a teacher that has access to future demonstration information, but the deployed policy does not require future observations. For robotics stacks that cannot rely on privileged future inputs at inference, this setup targets deployable policies while still leveraging hindsight during training.

    Signals: HINT-WAM does not improve every LIBERO-90 task: it lowers success on 22 tasks, with the largest reported drops including pick-and-place tasks into baskets or caddy compartments.

    Source: https://huggingface.co/papers/2611.24789

  • OpenAI Performance Factory

    TL;DR: OpenAI’s Performance Factory uses Codex-powered workflows to find and fix performance issues in ChatGPT.

    • The Performance Factory is a system of Codex-powered workflows that finds and fixes performance issues.

    Why it matters: Finding and fixing performance issues is intended to help keep ChatGPT fast.

    Source: https://openai.com/index/how-openai-performance-factory-keeps-chatgpt-fast

  • LangSmith normalizes coding-agent traces for cross-tool…

    TL;DR: LangSmith brings coding-agent sessions into a shared trace model for cross-tool cost comparisons, session analysis, and spend caps.

    • Integrations cover Claude Code, Codex, Cursor, GitHub Copilot Chat, Pi, and OpenCode; traces share fields for sessions, turns, tools, and metadata.
    • Filter by thread_id, model, provider, or tool name; compare token use, session cost, tool calls, and subagent activity across agents.
    • LLM Gateway caps costs at user, team, and org levels; routing to open-source models is planned.

    Why it matters: When each coding tool emits different logs and cost breakdowns, comparing agent sessions becomes an accounting and attribution problem. A shared trace schema plus spend caps lets teams apply consistent budgeting and investigate cost drivers (tokens, tool calls, subagent activity) across heterogeneous coding-agent setups.

    Signals: LLM Gateway routing to open-source models is described as forthcoming.

    Source: https://www.langchain.com/blog/fix-your-coding-agent-bill

  • Gemini Robotics 2.5 expands dexterity and reports…

    TL;DR: Gemini Robotics 2.5 expands dexterity across 200 tasks, with 30–100% higher success and 20–30% faster completion than baselines.

    • At Mercedes-Benz, Apollo 2 autonomously sorted automotive parts into trays; Gemini Robotics ER tracked workflow while 2.5 controlled its hands and arms (96% success).
    • Inference-time hazard guidance raised safe task completion 33% around obstacles and 43% near people.
    • ASIMOV-Jailbreak has hundreds of disguised unsafe-request prompts across dozens of scenes; Gemini 4 Argon achieved 99.5% resistance vs. 82% for Gemini Robotics ER 2.

    Why it matters: The production-line pilot provides evidence that a system combining workflow reasoning and action control can perform automotive kitting autonomously in an active workplace. The task and safety results also give practitioners concrete measures for evaluating dexterity and hazard-aware completion.

    Signals: The task-evaluation approach could scale to many more than the 200 distinct tasks tested.

    Source: https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/

  • LangChain raises $300M Series C at a $3B valuation

    TL;DR: LangChain raises $300M Series C at a $3B valuation to advance its mission of enabling companies to own their intelligence.

    • LangSmith combines runtime, improvement, and governance tools for agent deployment, evaluation, automated fixes, and control of model, tool, and data access.
    • Over 7,000 organizations use LangSmith; Clay monitors hundreds of millions of monthly agent runs, and Workday reduced the time to deploy a new agent from five months to under a day.
    • CapitalG led the round, with participation from GIC, DFJ Growth, IVP, and Sapphire Ventures.

    Why it matters: For teams standardizing agent delivery, this highlights LangSmith as a consolidated surface for runtime operations plus governance over model/tool/data access, with a concrete enterprise case study (Workday) tied to deployment cycle time.

    Signals: LangChain says it will share several new LangSmith capabilities at a release event on November 20.

    Source: https://www.langchain.com/blog/series-c

  • AlphaProof Code proves Signal code with over six million…

    TL;DR: AlphaProof Code verifies Signal’s compiled cryptographic code against human-reviewed specs in four weeks with AI agents and Lean.

    • Experts and models translated security goals into human-reviewed specifications; Gemini Carbon agents generated over six million lines of Lean proofs.
    • Lean checked every proof; AlphaProof Code flagged one percent of the codebase for human review.
    • Proofs covered message protection before and after temporary device compromise; verification took four weeks.
    • With two previously fixed bugs reintroduced in isolated code, the system produced machine-checked proofs that the buggy versions were insecure.

    Why it matters: This workflow uses AI agents to generate large proof artifacts while keeping correctness anchored in two gates: human-reviewed specifications and Lean’s proof checker. Flagging a subset of code for review focuses expert attention where the automated pipeline identifies uncertainty or mismatch, rather than requiring exhaustive manual inspection.

    Signals: Scaling these techniques requires continued research.

    Source: https://deepmind.google/blog/introducing-alphaproof-code/

Evaluation, grounding & uncertainty

  • OpenAI measures coding-agent acceleration and reports…

    TL;DR: OpenAI measures how coding agents accelerate AI research and finds their performance remains uneven.

    • The work measures coding agents’ acceleration of AI research.
    • It reports that coding-agent performance remains uneven.
    • It examines what these findings signal about recursive self-improvement.

    Why it matters: If you are using coding agents to speed up research workflows, “uneven” performance implies you should expect variability across tasks and treat aggregate acceleration claims cautiously when designing evaluation and rollout criteria.

    Source: https://openai.com/index/research-acceleration-jagged-frontier

Subscribe

Get Applied Signals weekly newsletter in your inbox