NewCore launches security-first identities for AI agents after closing $66M seed funding round launches a security-first identity platform to secure and govern humans and autonomous AI agents after raising $66M. Outcome engineers should treat managed agent identities as a first-class control for authorization, audit trails, and lifecycle policies — build identity and governance into your agent stack early (Principle 15, 10).

Hermes Agent’s new async subagents take aim at the blocking-agent problem adds asynchronous subagents so delegated child agents run in the background, unblocking chats and enabling interactive multi-agent workflows. Use async delegation to avoid blocking user interactions, design progress-tracking artifacts, and surface partial results for long-running tasks in your orchestrator (Principle 09, 03).

datasette-agent 0.3a0 adds user-approved execute_write_sql and CLI approval options, letting agents safely modify databases from chat and the terminal. This provides a concrete human-in-the-loop pattern: gate state-changing actions with explicit approvals and record provenance so agents can act without breaking your data integrity (Principle 15, 03).

SGLang adds DFlash to push Qwen 3.5 397B-A17B inference up to 4.3x faster integrates speculative-drafting DFlash to speed Qwen 3.5 inference up to ~4.3× in benchmarks. Higher speculative throughput alters inference cost and latency tradeoffs for agent planners — revisit batching, speculative decoding, and token budgets when designing real-time agent loops (Principle 11, 12).

Sakana AI launches Marlin: ‘ultra deep research’ agent producing 100+ page reports in 8 hours runs eight-hour autonomous research loops that produce fully cited, long-form strategy reports. Long-horizon agents expose the need for chaining tools, citation/grounding, checkpoints, and monitoring — plan for interrupts, verifiable artifacts, and auditability for multi-hour agent runs (Principle 09, 16).