Overview
Anthropic pursued emergency court relief while funding a larger commercial partner channel, showing legal contraction and enterprise expansion at the same time. AWS placed deterministic authorization between agents and tools, and Atlassian tied a 10% workforce cut to AI investment. EnterpriseOps-Gym then measured why the control layer matters: the best tested agent completed only 37.4% of its stateful tasks.
Developments
1. Anthropic seeks a stay as the Pentagon rules out new talks
Anthropic asked the D.C. Circuit to stay the Pentagon's supply-chain-risk designation during judicial review, estimating hundreds of millions to multiple billions of dollars in possible 2026 revenue loss. Pentagon CTO Emil Michael separately said there was no chance of renewed negotiations.
The revenue range is Anthropic's estimate, and the court had not granted emergency relief. With talks closed, the next commercial branch depends on judicial timing: a stay could interrupt the contractor purge, while a denial would leave migration work running through the merits phase.
Sources: Reuters on the stay request · Reuters on the Pentagon's position
2. Anthropic commits $100 million to its partner channel
Anthropic committed an initial $100 million in 2026 to the Claude Partner Network for training, technical support, market development, and co-marketing. It also plans to expand its partner-facing team fivefold and introduced a Claude Certified Architect credential.
The program distributes implementation capacity through consultancies and service firms at the moment federal channels are narrowing. Its impact will appear in certified staff, funded deployments, customer retention, and partner concentration, not in a model benchmark or the size of the announced pool alone.
Sources: Anthropic
3. AWS evaluates every AgentCore tool request against Cedar policy
Amazon Bedrock AgentCore Policy evaluates each AI agent-to-tool request at the gateway against Cedar authorization rules. Policies can be written directly or generated from natural language, use default-deny semantics, and run independently of model reasoning.
AWS also provides a log-only mode before enforcement and analyzes generated policies for overly broad or restrictive rules. The boundary is deterministic only for the attributes it receives; incorrect identity claims, incomplete tool schemas, or a sensitive action hidden behind an overbroad tool can preserve risk outside the Cedar decision.
Sources: AWS
4. Atlassian cuts 10% of staff while shifting investment to AI
Atlassian announced a roughly 10% workforce reduction, about 1,600 employees, to self-fund more investment in AI and enterprise sales while accelerating profitability. The company said Rovo had passed five million monthly active users and acknowledged that AI changes both its skill mix and the number of roles in some areas.
Management explicitly rejected a simple “AI replaces people” account, making the announcement a capital-allocation decision instead of a measured automation result. The cut immediately concentrates delivery and support work among fewer employees; product output, service levels, margins, and rehiring provide separate tests of the claimed reinvestment case.
Sources: Atlassian · TechCrunch
5. EnterpriseOps-Gym finds planning and refusal failures
A March 13 preprint introduces a containerized benchmark with 164 database tables, 512 tools, and 1,150 expert-curated tasks across eight enterprise functions. Among 14 tested frontier models, Claude Opus 4.5 led with 37.4% task success.
Supplying oracle human plans improved results by 14 to 35 percentage points, while the best model refused only 53.9% of infeasible requests. The paired failures locate two separate risks: weak strategy reduces completion, and weak refusal can mutate persistent enterprise state even when successful completion was impossible.
Sources: EnterpriseOps-Gym preprint