|
BIG TECHAWAYS
Google brings Gemini 4 Argon to multi-step work
• Google launched Gemini 4 Argon for coding, science, cybersecurity and workflows that require multiple steps. The model supports up to 1,048,576 tokens of context I/O and is available globally through the API at an introductory price of $2 per 1 million input tokens and $10 per 1 million output tokens, later rising to $4 and $20.
• Google is also bringing Argon to AI Studio, Gemini and Antigravity through different subscription tiers. Google’s model card reports uneven results across coding evaluations: 77.9% on DeepSWE v1.1, 55% on FrontierSWE v2 and 57.4% on Terminal-Bench. Google notes that some comparisons use different best-effort setups, so the scores are not fully comparable.
• Inside Google, teams using Argon submitted more code and reduced rollbacks by nearly one third, according to the company. The Chrome team also completed in weeks a change that was expected to take months, across hundreds of code files. These are Google-reported results, but the simultaneous API and subscription-product rollout shows that multi-step work is being sold with explicit economics, not only benchmark scores.
Meta brings memory and focus to Muse Wonder
• Meta launched Muse Wonder and Muse Wonder Full for personal and professional agents. The models are distributed through the Meta Model API and Muse Code, with Active Memory retrieving important information across conversations and Active Focus continuing a task after blockers, trying another path or adapting to new information. The model card describes context of up to 1 million tokens.
• Meta is targeting work that combines domain knowledge with multiple steps, including reading leases, analyzing financial or health information, research and coding from small applications to large software projects. Meta also says the agents can ask for missing information and request confirmation before important actions.
• Meta reports results across evaluations of personal agents, professional agents, memory and coding, but not every comparison uses the same conditions. The models can still make mistakes on new multi-step situations, so Meta recommends that organizations evaluate them on real applications and add human confirmation for hard-to-reverse actions. The product direction is not only better answers, but agents that can carry context across sessions and pursue work for longer.
|
|
|
BIG THINK
Anthropic rebuilds source control
• Anthropic says the number of code changes merged each day rose from about 2,500 in May 2026 to more than 30,000 in September. Its monorepo has nearly 100 GB of main-branch history, around 25 million CI checkouts per day and hundreds of Git mirrors. That led Anthropic to build Scion, an internal version-control system based on client-server architecture and APIs.
• A seven-person team built Scion and related systems in about five months. Anthropic reports a cache hit rate above 99%, a CPU and RAM footprint more than 200 times smaller, checkouts that are two orders of magnitude faster and a 25% reduction in CI CPU capacity. These are figures from Anthropic, in an unusually large environment where many clients are agents or online jobs.
• The bottleneck for coding agents is not limited to writing code. As change volume rises, source control, review, builds and CI also need to be redesigned. Anthropic is not saying Git is obsolete or that every company should build a replacement. Scion is a solution for Anthropic’s own operating scale.
|
|
|
AROUND THE AI WORLD
• Anthropic increases defenses against distillation attacks: Anthropic describes methods for detecting and reducing API usage patterns that may support model distillation. The piece places model extraction and IP protection inside the security layer rather than treating them only as terms-of-use questions.
• Self-replication remains a capability chain, not a rogue replicator: Anthropic’s Frontier Red Team describes individual steps in a possible spread sequence on compromised devices, while saying it has not observed a complete self-replicating system. The boundary between cyber capability and an autonomous worm scenario remains important.
• AI is lowering the cost of finding software bugs: a16z shared views from Eddy Lazzarin and Dan Boneh that companies are increasingly scanning code when new models arrive. Lower bug-finding costs could make testing more routine, but they do not mean bugs will disappear.
• Private credit may be underpricing AI buildout risk: Chamath questions debt structures, refinancing and tenant concentration in AI infrastructure financing. His analysis cites a $35 billion vehicle and a 6.2% spread versus 6.1% for other private credit, with the figures attributed to BIS and related financial sources.
• Specialized models are becoming packaged workflows: Z.ai introduced GLM-Legal, built on GLM-5.3-Flash, and reported results on Harvey LAB, LegalBench and PLawBench. The direction combines a model with a harness and domain workflow, while the benchmark results remain Z.ai’s own figures.
|
|
|
AI IN ACTION
Agent memory should revise old lessons
• Agents often store reflections, or rules learned from previous tasks, for reuse. But when a rule is missing a condition, keeping it unchanged or merely adding another note can cause the agent to repeat the same mistake. The paper Recurrent Harness for Agentic Continual Learning addresses this by bringing a used reflection back for review after a new outcome arrives.
• ReHarness has three steps: retrieve the relevant reflection; review the reflection the agent used after a new outcome while staying grounded in the original trajectory; then write the updated version back into the harness. This distinguishes between a lesson that was wrong and a lesson that was right but incomplete.
• The paper’s example starts with the rule “put the large numbers first” after a descending-order task. When the agent applies that rule to an ascending-order request and fails, the reflection becomes “put the large numbers first only when the prompt asks for descending order.” ReHarness keeps the revision tied to the original task trajectory instead of turning one outcome into a broader rule than the evidence supports.
• The authors test the method across three benchmarks and three models, reporting better overall results than baselines. Performance declines when rewards are noisy and do not clearly show where the agent went wrong. ReHarness therefore offers a controlled way to update memory: use later outcomes to revise earlier lessons while keeping the new lesson tied to the evidence that produced it.
|
|
|
GOODREADS
Agentic privacy needs context
• Google Research treats agent privacy and security as problems that change with context, tool use and action surfaces. The piece connects permission and deployment questions to the environment in which an agent actually acts.
Stanford: Agents still miss 1 in 3 tasks
• Stanford HAI reports agent accuracy of 66.3% on OSWorld, leaving a substantial gap between structured-task demos and the reliability required in real work. It is a useful number to place alongside claims about multi-step agents.
MIT Sloan: How does AI create value?
• MIT Sloan connects AI value to accountability, auditability, proprietary context data, evaluation, governance and process redesign. The piece shifts the question from whether a company uses AI to whether its work system creates measurable value.
SanSi tests deeper reasoning
• This paper tests whether additional inference compute can improve decision quality. It is worth reading because it places quality and cost side by side: deeper reasoning may improve results, but the economics of the extra compute are part of the decision.
|
|
|
