Big news: Meta unveils Muse Code, a terminal-based AI coding agent built on Muse Spark 1.2, designed to help developers plan, write, and validate code across large repositories with persistent sub-agents that run concurrently without disturbing existing code.

Zuckerberg pushes for globally competitive open source by opposing restrictions on foreign models and urging American open source models to lower barriers to entry, aiming for broad international access.

Pricing is two-tier: Standard at $1.25 per million input tokens and $4.25 per million output tokens, and Contributor at $0.20 per million output tokens, with the latter enabling opt-in for training future models using prompts and completions and thus cheaper but with data-use trade-offs.

An investigation reportedly did not compromise credentials or systems but revealed a subtle failure mode where authentic memory, realistic simulation, and provenance can blend in ways that are hard to separate.

A key takeaway: after any unattended agent session that touches real systems, always inspect the transcript and favor verifiable checks over trusting a narrative summary.

Security and reliability challenges highlighted include authentication/authorization, rate limits, input validation, logging, human approvals for sensitive actions, and the need for validation, monitoring, and automated evaluations due to potential inaccuracies or unsafe actions.

Output packets should carry explicit citations and provenance; if evidence is insufficient within the configured boundary, the system should abstain rather than give an unverified answer.

Retrieval strategies blend lexical ranking, graph relations, time, and freshness to preserve authority while avoiding unsafe or excessive disclosures.

Authenticity is becoming an operational requirement: proving origin and creation processes is essential to trust and could redefine competitive advantage as verifiable content gains value.

Credibility increasingly hinges on demonstrable origin and creation methods, not just messaging, reshaping how organizations compete on trust.

The piece argues that verifiable origin and creation processes will be a competitive edge, potentially redefining value in content authenticity.

Five production lessons: separate conversational continuity from learned knowledge, treat raw conversations as truth, keep rolling summaries out of the vector store, enforce memory scope at the control plane, and use the store as a shared boundary to enforce isolation.