01
Your procedures are now portable files. Vet them like an authority you did not look up yourself.
The agent procedure format has spread across vendors, with OpenAI Codex, GitHub Copilot, Cursor and Microsoft's agent framework among more than 30 tools reading the same file. A working method written once can be dropped into another tool, another team, another firm.
CaseFlow angle. Treat every shared or downloaded procedure the way you treat an authority somebody else cited. Record who wrote it, when it was last tested, which matter types it is cleared for, and who signed it off for live files. No name against it, no run on a client matter.
02
Audits found flaws in more than a third of published procedures. Read them like code, not like guidance.
An audit of 3,984 published skills found 36.8% carried at least one security flaw, 13.4% carried a critical one, and 76 contained confirmed malicious payloads. A wider analysis of 42,447 skills put the vulnerability rate at 26.1%. Benchmarking also found narrow procedures beat one large catch-all.
CaseFlow angle. Nothing downloaded goes near a live matter until someone has read it line by line and run it on a closed file first. Split the big catch-all procedure into narrow ones per document type. If nobody in the firm can read it, do not run it.
03
The cheapest strong model yet arrived chattier. Re-run your templates before you trust the output.
Claude Opus 5 landed as Anthropic's strongest and most cost-effective flagship and is now the default in Claude Code. Early feedback flags backwards-compatibility friction and noticeably longer, more verbose outputs.
CaseFlow angle. A model swap under your supplier's product can change the length, tone and structure of a templated document without changing the button your handler presses. Ask each supplier for the date they switch, then re-run a sample of your five most-used document types and compare against the version your team actually signed off.
04
You can now see what each automation costs per matter. Start pricing it properly.
Anthropic's July release notes added programmatic usage, cost and adoption analytics down to individual saved procedures. Administrators can see which automations are actually used and where the model spend lands.
CaseFlow angle. Pull the usage figures for your three most-used AI procedures and divide by matters opened. That number is your AI cost per file, and it belongs in front of whoever sets your fee scales. Retire anything nobody has run in ninety days.
05
Who is liable when the software was driving? Motor claims is not ready for the answer.
Intact's technical motor lead has warned that insurers are unprepared for software liability and recovery on connected and autonomous vehicles. The legal questions are arriving well ahead of the process that has to handle them.
CaseFlow angle. Add two fields to your first notification script now. Was driver assistance engaged, and does the vehicle hold telematics or event data. Liability and recovery arguments on these files will turn on data that is gone within weeks if nobody asks for it on day one.