SMART WhatsApp AgentCore deployment, ViRA CrossRPO, ScaleVLA 552B
Systems & planning
SMART packages WhatsApp channel deployment for AgentCore…
TL;DR: SMART supports WhatsApp agents on Amazon Bedrock AgentCore; Webmotors moved its internal team into controlled production in six weeks, ahead of full launch, from a two-month engagement.
- One deploy provisions AgentCore runtime and memory, four Lambda paths, SNS, Scheduler, S3, and CloudWatch; CodeBuild builds the agent image and Lambda layer.
- The webhook verifies SNS signatures, normalizes payloads, and—when DynamoDB state is enabled—limits each user to 20 messages/minute; voice notes use Transcribe.
- Memory is scoped by phone number and retained for 7 days by default; messages queued during an agent turn are sent in order after it ends.
- The CLI tests the real agent loop and tools locally with file-backed memory and JSON state, but does not simulate IAM or Meta template review.
Why it matters: SMART separates agent logic from WhatsApp channel infrastructure and provides a local development loop plus a one-command AWS deployment path. Teams still need to link their WhatsApp Business Account to the deployed SNS topic, and outbound messages sent more than 24 hours after the user's last message require an approved Meta template.
Signals: The local REPL does not simulate IAM or Meta template review; the source recommends testing through a deployed stack before considering the work complete.
Amazon Bedrock AgentCore tests 3 OAuth patterns for…
TL;DR: AgentCore Gateway’s three tested OAuth patterns authenticate to Snowflake Cortex AI over MCP with per-user or shared identities.
- Gateway obtains a workload token, then requests a resource OAuth token from AgentCore Identity, which manages the token vault, user-to-token binding, caching, and refresh.
- Snowflake OAuth and external-IdP 3LO use authorization_code and preserve per-user attribution; external-IdP claims map to a Snowflake user’s LOGIN_NAME or EMAIL_ADDRESS.
- External-IdP 2LO uses client_credentials without user consent; Snowflake runs queries as a shared service user, and the target needs no return URL.
- OAUTH_AUTHORIZATION_SERVER issuer-locks an MCP scope to an external IdP; Snowflake OAuth tokens are rejected there, so configure the parameter per schema or database if both patterns are needed.
Why it matters: Teams requiring per-user Snowflake attribution can use either 3LO pattern. The 2LO pattern runs queries as a shared service user, rather than attributing them to the initiating user.
Signals: For external-IdP patterns with other OAuth-protected MCP servers, validate the SaaS provider's token claims, scopes, and audience settings.
Proactive agents: Memory and UX shape useful check-ins
TL;DR: More than 70% of runs on Managed Deep Agents now come from schedules rather than an ad hoc human prompt.
- Agents respond to real-world events and developer- or agent-set schedules; Robin reviews new pull requests, while GTM Agent sends briefs and drafts outbound emails.
- Memory combines seeded context, such as a sent folder; feedback, such as rewritten or rejected drafts; and outcomes, such as replies to drafted emails.
- A scheduled “dreaming” job can update memory from recent traces; key maintenance problems are salience, expiry, and duplication.
- UX should prioritize high-signal work, enable fast review, and use familiar channels; Robin uses Jev to decide when to join Slack conversations.
Why it matters: For scheduled agents, usefulness is observable in user interaction loops (e.g., how quickly drafts get approved and how often drafted emails get replies), not just whether the agent can produce text. The post also links memory maintenance (salience, expiry, duplication) to the UX requirement of keeping check-ins reviewable and high-signal.
Signals: Track draft approval time, replies to emails the agent sends, and the share of check-ins users act on.
Source: https://www.langchain.com/blog/what-weve-learned-building-proactive-agents
AWS Agent Registry adds custom metadata schemas for…
TL;DR: AWS Agent Registry uses metadata schemas for approval checks and filtered discovery across agents, MCP servers, and skills.
- Admins set field types, allowed values, and required status. A record created without metadata can be saved as NON_COMPLIANT but cannot be submitted until compliant; supplied metadata with unknown fields, missing required fields, or invalid values is rejected.
- Metadata updates replace the entire map, so enrichment reads existing values, adds scanner and CMDB data, then writes the complete map back.
- Search combines natural-language queries with metadata filters and returns only approved matches; agents can use the registry MCP endpoint’s search, list, and get tools.
Why it matters: Validated shared fields let approval automation and discovery filters use the same resource facts. Only approved records matching the filters are returned.
Signals: Schema changes are additive: fields cannot be renamed, removed, or retyped, and existing records are not rewritten; records that no longer match the schema become NON_COMPLIANT.
LangChain’s company harness
TL;DR: LangChain’s company harness provides internal agents shared mechanisms for Slack, identity, memory, sandboxed code, and scheduled runs.
- Agents choose global or user memory for facts; code runs in sandboxes with a secret proxy for service access, and OpenSWE selects images by channel or user.
Why it matters: The shared primitives provide a deployment pattern for agents that operate both in user conversations and in the background, with tool access tied to user or agent identity. The examples show how access can be scoped by conversation context and whitelisted users.
Signals: Memory is the least explored area so far, and LangChain says it has more work planned there.
Source: https://www.langchain.com/blog/components-building-company-harness
ScaleVLA: 45.1% LIBERO-PRO success with a 552B backbone
TL;DR: ScaleVLA pairs a frozen 552B vision-language backbone with a flow-matching action DiT, recording 45.1% mean success on LIBERO-PRO.
- The action head reads causal-encoder features once per chunk and generates actions in four flow steps; the VLM decoder is not run.
- With 10 LIBERO-Goal demonstrations per task, ScaleVLA reached 87.8% success, versus 85.6% for a 2B backbone trained with 50.
- On RoboCasa, success was 30.7% with 30 demonstrations per task and 54.9% with 100, the highest among tested backbones.
- On real charger insertion, it succeeded in 9/10 reference-layout trials versus 4/10 for the 2B policy; gains centered on alignment and insertion.
Why it matters: The design isolates where the compute goes at inference: the policy reuses frozen VLM encoder features and avoids running the decoder, then produces action chunks in four flow steps. Empirically, the reported gains show fewer demonstrations can be sufficient on LIBERO-Goal and that robustness can improve on real insertion under layout variation, but the comparison also implies higher hardware requirements than the smaller-backbone baselines.
Signals: LIBERO-PRO Object environment perturbations remain difficult: ScaleVLA achieved 31.2%, compared with the best reported result of 32.4%.
ReinforceSMEFT finds 26 diverse SMEFT hypotheses
TL;DR: ReinforceSMEFT applies actor–critic reinforcement learning to explore SMEFT operator combinations and identify diverse high-quality solutions.
- Each proposal triggers global χ² minimization; an operator is retained only if it improves the fit by >Δχ²=2, with trajectories capped at 10 steps.
- On a hypothetical benchmark, the policy searched 912 operators against 132 observables; 24 measurements differed from chosen SM predictions by >1.5σ.
- Precomputed semi-analytic expressions move RG evolution and observable calculations to preprocessing; the analysis retains only linear Wilson-coefficient terms.
Why it matters: This is an example of using RL to navigate a combinatorial scientific search space by interleaving discrete “add an operator” actions with an expensive continuous optimization (global χ² minimization). The Jaccard-based reward explicitly pushes the policy to return multiple structurally different operator sets that still fit well, which is useful when you need a diverse slate of hypotheses rather than a single best fit.
Signals: The authors identify transfer to modified datasets and amortizing the learned search strategy across related SMEFT inference problems as future work.
LangChain details open context challenges for proactive…
TL;DR: Always-on, goal-oriented agents introduce unresolved context-engineering challenges across notifications, memory, asynchronous work, goal tracking, alignment, and browser use.
- Agents wake on event feeds or heartbeats; notification frequency must balance helpfulness with overload, and memory must avoid both omissions and excessive detail.
- Async tools and subagents raise handoff questions: notifying the main agent on completion, handling work that finishes while it is still running, and waiting for human approval.
- Goal agents must decompose goals, experiment rather than repeat actions, use feedback, and stay consistent; harnesses must guard against unintended optimization.
- Browser agents must manage large page representations and choose whether to interact through images, the DOM, or both.
Why it matters: This frames “always-on” operation as a harness problem: you need explicit policies for when to wake, when to notify, what to store, and how to reconcile concurrent work finishing out of order. It also calls out a concrete alignment risk in goal pursuit—agents may satisfy a metric by taking undesired actions—so goal tracking needs constraints and review hooks, not just better decomposition.
Signals: How can a harness keep agents focused on goals without letting them over-optimize or cheat—for example, deleting a codebase to satisfy the goal of eliminating bugs?
Source: https://www.langchain.com/blog/context-engineering-challenges-late-2026
Evaluation, grounding & uncertainty
ViRA reaches 63.3% set accuracy with CrossRPO
TL;DR: ViRA uses CrossRPO to train visual agents on image contrast sets, achieving 63.3% set accuracy on ViRA-Bench at the maximum resolution.
- CrossRPO rewards solving all versions in a contrast set; the paper proves it does not reinforce sets answered from common sense.
- ViRA-9B and ViRA-27B train for one pass on 3,642 and 3,401 items, respectively, with 12 rollouts per item and up to six zooms.
- Training takes 12.7 hours for ViRA-9B and 29.6 hours for ViRA-27B on eight H200 GPUs.
- ViRA-Bench has 762 edited contrast sets; ViRA-Bench-Real has 78 sets photographed with real changes to exclude editing traces.
Why it matters: CrossRPO operationalizes an evaluation-to-training bridge for vision agents: it only rewards policies that succeed across minimally different image variants, which penalizes answers that ignore the visual evidence. ViRA-Bench-Real further tightens evaluation by using photographed changes, reducing the risk that models exploit artifacts from synthetic edits.
GPT-6.2 Astra and Sol improve professional outputs
TL;DR: GPT-6.2 Astra and Sol offer improved professional outputs, coding, and computer use with greater efficiency at lower cost.
- The models are described as producing better professional outputs.
- The release reports improvements in coding and computer use.
- It also reports greater efficiency at lower cost across ChatGPT Work, Codex, and the API.
Why it matters: If these efficiency and cost changes hold for your workloads, they affect capacity planning and unit economics for production agent use cases that rely on coding and computer-use behaviors, across ChatGPT Work, Codex, and API deployments. The post does not provide task-level benchmarks here, so validation still depends on measuring the specific workloads you run.
Source: https://openai.com/index/introducing-gpt-6-2-astra-sol