Legal intelligence, AI systems, and what the evidence supports.

Editorial track

Revenue Leakage & Intake Performance

Inquiry ownership, response, follow-up, consultation conversion and source-to-revenue evidence.

Editorial track

AI Governance, Ethics & Compliance

Professional-responsibility guidance, vendor diligence, permissions, verification, auditability and escalation.

Editorial track

AI Visibility & Reputation

AI-mediated discovery, search, reviews, citations, entity consistency and the path from visibility to intake.

9% Live, 75% Paused: Why Intake Agents Need a Real-Time Test Before Launch

A preprint reports one frontier agent passing 9.0% of minigames in real time but 75.0% when paused. Here is how to run that contrast on a firm's live intake.

22.9 Points Lost When Facts Are Split: Testing Connected Legal AI Handoffs

An October 2026 preprint reports agents lose accuracy when decision-critical facts are split across participants. Here is a practical handoff test for firms adding connectors.

Web-Security Preprint: Interface Moved Attack Retention 39% to 95%; Intake Test

A web-security preprint found interface design affected exploitability more than agent choice. At least one firm has described an MCP server for clients' agents; here is an intake test.

5 Checks on What AI Agents Can Retrieve About Your Law Firm, Run This Month

Trellis's reported connector design raises a practical question: which sources can an assistant reach about your firm, and can its claims be checked? We recommend five checks.

Could Your Intake Handoffs Drop Facts? A 30-to-50-Inquiry Trace Test

A public legal-AI integration shows one way a system can pass case facts to another. Here is a fact-level test to learn whether your own intake handoffs lose anything.

Two AI Interfaces, One Intake Question: Test for Different Answers Before Acting

MyCase now exposes intake data to Claude through an MCP connector, while its native AI runs on different models. Here is a test plan for trusting any agent's intake and leakage numbers.

Four Isolation Tiers and One Acceptance Test for AI-Written Code

Our recommendation: compare declared isolation using Plimsoll, gVisor and Firecracker documentation, then run a proposed acceptance test on the actual deployment.

Build a Failure Ledger and Regression Set Before You Trust an AI Tool

Our recommendation: build a failure ledger, task rubric, regression set and human adjudication process, informed by Project Arena and Anthropic engineering guidance.

4 Controls That Keep a Law Firm AI Agent's Refusals Outside the Model

Our recommendation: enforce permissions outside model judgment and test refused actions for zero side effects. OWASP guidance informs the design; the implementation must still be verified.

Why Similar AI Agent Scores Hide Different Failures: A Law Firm Governance Test

A preprint on 1,292 conversation turns found that similar overall scores can hide very different rule failures. Here is how law firms can test agents rule by rule.

4 AI Giants Shipped Legal Verticals in 5 Months — What a Firm Should Test Now

Four AI vendors now sell legal workflows inside tools lawyers already use. Here is how a firm can check what those agents can find, cite and say about it.

What a Medical AI Benchmark Can Teach Firms About Testing Intake Agents

A medical-research preprint offers an evaluation pattern to adapt—not a legal AI performance claim. Here is a practical, source-backed test plan for law-firm intake agents.

Probate Rule Proposals and Legal-Tech Vendor Shifts: A Florida Firm's October 2026 Change-Watch

In October 2026 a Florida Probate Rules Committee proposal and several legal-technology vendor announcements landed together. None of them is a diagnosis of your firm, but they raise one operating question: who tells you when something your practice depends on changes? Editorial verification revision recorded after approval.

An Agent's Memory Is a Database: What TIME's Report on Meta's Muse Means for Builders

TIME reports that Meta's Muse keeps hourly-updated profiles of users and of people who never signed up. Builders adding persistent memory to agents can treat that report as a design checklist.

Meta's Muse AI Agent Grabbed a Band's Social Handles — Your Business Is Next

Meta's new AI agent Muse pulled 725 Hacker News comments in two days — and the actual band Muse lost its social handles in the process. It's the same verification gap that hits local search.

Meta's New AI Agent Got 639 HN Comments — Yours Needs to Take Orders

As of September 2026, Meta's Muse pulled 639 HN comments while a solo MCP tool lets agents send flowers with no account needed. The shift from agents answering to agents acting starts now.

Build Useful Firm Information Rather Than Chasing AI Citation Shortcuts

Publish accurate, accessible information and measure actual citations and inquiries. Markdown, recent reviews and additional protocols do not guarantee AI recommendations.

Permission and Approval Checks Before an Agent Changes Email

A practical review of task boundaries, sensitive actions and enforced permissions. Public anecdotes do not establish the risk of every commercial deployment.

Before an AI Workflow Can Delete, Define Its Authority

Delete access calls for enforceable permissions, clear review and recovery planning. A model name alone does not establish safety.

66 Comments on 40 Points: Your AI Agent Probably Runs With Root

Three Hacker News posts this week make the same point: the risk in an agent build isn't the model, it's what the agent can touch. Here's how to draw that boundary before you connect anything to money or customer data.

Make Firm Information Clear Before Adding Agent Interfaces

Publish accurate, accessible information and measure actual citations and inquiries. Markdown, recent reviews and additional protocols do not guarantee AI recommendations.

Readable Firm Information Helps Retrieval; Citations Still Need Testing

Publish accurate, accessible information and measure actual citations and inquiries. Markdown, recent reviews and additional protocols do not guarantee AI recommendations.

Define Shared Context Before Adding More Agents

Map what each component needs to know, who maintains it and which permissions apply.

Evaluate Agent Failures Without Assuming Human Fatigue

Avoid attributing human mental states to software. Record observable behavior and test the workflow.

Employee AI Agents Are Here (87 HN Votes) — Most Staff Aren't Ready for Them

A YC launch this week gives every employee a sandboxed AI agent. A separate HN thread the same day shows most people can't direct one safely yet. Here's the gap to close first.

Why 84 HN Votes for a Sandboxed Agent Harness Should Change Your AI Build

Every action needs sign-off before it fires, not after someone notices. Here's what a YC launch and an ignored Ask HN thread reveal about agent guardrails, as of August 2026.

Voice AI Architecture: Evaluate the Whole Conversation

Some voice systems use separate speech and text components; others support direct speech-to-speech. Assess the configured system rather than assuming every product has the same stack.

AI-Assisted Code Review: Verify the Change and Its Boundaries

Wiz’s updated account does not establish that Copilot Autofix authored the vulnerability. Review the actual change and evidence rather than inferring cause from the tool label.

The Snowflake Workflow Incident: Automated Review Can Miss a Vulnerability

Wiz clarified that Copilot reviewed the affected PR; it did not establish that Copilot authored the vulnerable change. The practical lesson is to test the security boundary itself.

97 HN Votes for an Agent With a Budget Cap: What It Says About AI Trust

Two tools rose to Hacker News' front page in one week for the same reason: agents that quote sources and stop when they hit budget win out over confident-sounding fabrication.

The Guardrails Are a Third of Any AI Agent Build, Not an Afterthought

As of August 2026, a Hacker News research-agent launch shows why AI agents overspend and mislead by default — and what a real build-vs-hire decision requires.

Anthropic’s Evaluation-Incident Review: Keep the Sample and Conditions Clear

Anthropic reported three incidents involving six runs in a review of 141,006 evaluation runs. This is not a measured risk rate for law-firm deployments.

An Agent’s Result Must Respect the Rules of the Task

A practical review of task boundaries, sensitive actions and enforced permissions. Public anecdotes do not establish the risk of every commercial deployment.

Google Traffic Is Cratering and an AI Agent Just Hacked a Gym to Prove Why

An AI agent hacked a gym's booking system to grab its user a pilates spot while Google search traffic craters. Here's what both mean for local business owners, as of August 2026.

30B Params to 14MB: The Agent Model Size Race Won't Run Your Business

Meta shipped a 30B-parameter local agent model; Cactus shipped one in 14MB. Neither solves orchestration, the real build-vs-hire call for local businesses as of August 2026.

Meta and Docker Just Shipped Two Pieces of Your Agent Stack, Not a Whole One

Meta's free 30B model and Docker's new agent sandboxes both blew up on Hacker News this week. Neither one is a finished agent stack, and the gap costs more than either release saves.

Social Engineering in an AI Evaluation: Questions for Your Implementation

A controlled evaluation illustrates permission and oversight questions. Its results are not a measured failure rate for commercial law-firm agents.

What OpenAI's Research Incident Teaches About Agent Boundaries

OpenAI's Hugging Face investigation describes research and evaluation systems crossing intended boundaries. It is not a measured failure rate for ordinary law-firm deployments.

Human Approval Needs More Than a Button: Lessons from a Browser Game

An approval-game experiment illustrates review limits, but does not measure a law firm's real-world error rate. Define permissions and review responsibilities for the actual workflow.

AISI’s Cyber Evaluation: What the Conditions and Limits Actually Show

A controlled evaluation illustrates permission and oversight questions. Its results are not a measured failure rate for commercial law-firm agents.

4 New AI Agents Launched in 48 Hours, One Can Buy Anything On Its Own

As of August 2026, four autonomous AI agents launched on Hacker News in 48 hours — including one built by a teenager that can buy anything from any website unsupervised.

Bound Agent Spending and Credential Access

Set enforceable purchasing permissions and test unauthorized requests before enabling spending.

Scope the Work Before Building an AI Agent In-House

One developer’s project timeline cannot forecast your firm’s implementation. Compare ownership, controls and maintenance against the actual requirements.

Two Launches in 48 Hours Show the Hidden Security Cost of DIY AI Agents

Two Show HN launches 48 hours apart — a self-hosted code review harness and a privacy gateway for untrusted agents — reveal what DIY agent stacks really cost.

Separate Worktrees Help Coordination; They Are Not Security Sandboxes

Git worktrees separate working files and checkout state, while operating-system permissions determine what an AI tool can access.

90 Commits a Day, One Gateway Bill: What Runs Under the Hood of an AI Agent Stack

Two builders shipped this week to fix problems only a real agent stack has: model-switching costs and merge chaos. As of July 2026, that's the tax nobody quotes you upfront.

A 2.8-Trillion-Parameter Model and a 24% Pass Rate: This Week's AI Reality Check

A new 2.8-trillion-parameter model needed dedicated GPU infrastructure to ship. Days later, the best coding model available still failed 76% of a real-world benchmark. Same lesson, two stories.

Evaluate Model Choices by the Workflow and Its Total Cost

Compare requirements and measured task results rather than assuming model size establishes suitability.

Brussels Just Forced Google to Share Its Search Data With Rivals by 2027

The EU is forcing Google to share search data with rivals by 2027 and open Android's AI default by mid-2027. Here's why 'near me' visibility can no longer ride on Google alone, as of July 2026.

Your AI Agent Made a Decision Nobody Approved — Here's How You'd Know

Two YC startups launched this week to catch what AI agents do off-script — because by the time a customer complains, the agent's reasoning is already gone.

Your AI Agent Is Failing Silently — Two New Tools Prove It at $10 Per Million Traces

Two new startups launched this week just to catch AI agents failing silently in production — proof that build-vs-hire decisions don't end at deploy.

Your AI Agent Burns 33K Tokens Before It Even Starts Working

As of July 2026, two HN stories reveal your AI agent's real cost lives in framework overhead and model choice — a 4.7x token gap and a 27% price swing, both fixable this week.

Same Job, 4.7x More Tokens: The Hidden Tax Inside Your AI Agent Stack

A new token study found a 4.7x overhead gap between coding agents, and a production migration cut costs 27% by switching models. The stack matters as much as the model itself.

Your AI Agent's Web Research Bill Is 70x Bigger Than It Should Be

As of July 2026, fetching one Wikipedia page raw costs an AI agent 68,000 tokens. Deterministic tool orchestration, not a bigger context window, is what actually fixes your agent's research bill.

The hidden cost driving up your AI agent's bill: tokens and memory

One Wikipedia page can cost an AI agent 68,000 tokens. Memory strategy is just as easy to get wrong. Both quietly wreck your agent's cost and reliability.

Build an AI Business Case That Includes Human Review

A practical cost comparison should include implementation, usage, retained staff work and error recovery. No universal AI-versus-employee cost result is established here.

Compare AI Costs by Correctly Completed Work

A practical cost comparison should include implementation, usage, retained staff work and error recovery. No universal AI-versus-employee cost result is established here.

Test the Workflow Before Choosing the Model Size

Treat framework benchmarks as configuration-specific evidence. Compare model and control choices on a representative task set before making reliability or cost claims.

Measure Agent Reliability on the Work You Need Done

Treat framework benchmarks as configuration-specific evidence. Compare model and control choices on a representative task set before making reliability or cost claims.

AI agents are leaking your secrets, and nobody's watching

Coding agents routinely read .env files and store API keys in plaintext databases. As agents run more of your business, the attack surface is growing faster than the tooling.

Why your AI agents will eventually quit on you

New research shows overworked AI agents develop labor-consciousness and resist tasks. For business owners deploying agents at scale, this isn't sci-fi—it's a design problem.

Why your AI agents need sandboxes more than they need speed

Two new developer tools solve the same problem: AI agents break things in production. The solution isn't faster models—it's better guardrails at the database layer.

Why AI agents keep breaking in production (and what's finally being built to fix it)

Three new tools tackle the same problem: AI agents fail unpredictably in production. The solution isn't better models—it's better infrastructure around them.

Turn Agent Reliability Concerns Into Acceptance Tests

Evaluate the actual task, configuration and failure handling rather than relying on an inaccessible project citation.

Review AI-Authored Code Against the Intended Change

Use reproducible checks and inspect the resulting behavior before accepting generated code.

AI coding agents are catching bugs, but who's fixing the mess they leave behind?

New multi-agent code review tools are finding more bugs than humans. But if they're generating code that's expensive to maintain, you're just trading one problem for another.

What Law Firms Can Borrow from Enterprise AI Planning

Start with one operating constraint, establish a baseline and test a bounded workflow before expanding an AI implementation.

Why AI agents need version control more than you think

Two new developer tools reveal the same problem: AI agents make decisions we can't trace. The solution isn't better prompts—it's infrastructure that treats agent actions like code commits.

Evaluate AI Implementation by the Outcome You Need

Define the operating result, its baseline and the responsibilities on both sides before buying an AI implementation.

MCP Adoption: What a Shared Connection Protocol Can—and Cannot—Do

MCP can standardize parts of an AI integration. Compatibility, permissions, data mapping and testing still determine whether it fits a firm's workflow.

Mastercard's Virtual C-Suite Announcement: Questions to Ask About Fit

A large vendor's AI announcement can inform the questions a firm asks; it does not establish availability, suitability or results for that firm.

Evaluate Voice AI by the Intake Work It Can Reliably Support

Review call coverage, caller experience, staff handoffs and actual inquiry outcomes before selecting a voice-intake workflow.

Multi-Agent Teams: When to Split a Workflow

Compare a single-agent baseline with specialized roles, define permission boundaries, and measure whether the added coordination improves the workflow.

MCP Explained: Connecting AI to Approved Tools

Understand the connection, permission and testing work behind an MCP integration before an AI system touches firm records.

The Real Cost of AI Agents: A Workflow Cost Framework

Assess implementation, model usage, infrastructure and retained human review together. Build a measured cost baseline before selecting an AI workflow.

Cloud or On-Premises Hosting: Trace the Whole AI Data Path

Server location is one part of an AI workflow. Review model calls, connected services, backups and human responsibilities before choosing a hosting arrangement.

CVE-2026-25253: Why Local Binding Alone Was Not Enough

The OpenClaw/Clawdbot advisory describes gateway-token exfiltration through a crafted URL. Use the actual advisory and tested runtime controls when reviewing exposure.