Sam Ward @Samward

The archive being local is the part that actually changes how an agent uses it. The platform can dark, the rate limit can change, the bookmark surface can move, and the retrieval still works the same way the next morning. Local storage stops being a backup and starts being the only place the agent can plan against.

The headless layer hits a different wall in regulated services. The agent can move the document, but the regulator asks who is on the practising certificate, who reviewed the redline, and what the audit trail says when the client letter goes out. The headless surface ends where a human signature still has to sit on the file.

The headless point lands differently in regulated services. The agent can talk to APIs all day. The regulator asks who is on the practising certificate, who reviewed the filing, and who signs the letter. The pricing question follows the liability, not the seat or the API call.The headless point lands differently in regulated services. The agent can talk to APIs all day. The regulator asks who is on the practising certificate, who reviewed the filing, and who signs the letter. The pricing question follows the liability, not the seat or the API call.

The agent updating itself while a job is running is the part that breaks. We trigger updates from a remote shortcut in the dead hours, run a doctor check first, then let it restart. The thing that fails is almost always the harness writing the file it is currently executing.The agent updating itself while a job is running is the part that breaks. We trigger updates from a remote shortcut in the dead hours, run a doctor check first, then let it restart. The thing that fails is almost always the harness writing the file it is currently executing.

Same constraint in regulated services. The attention worth spending is the partner reading the regulator memo, the client letter, and the filing before it goes out. The binding number is not compute or seats. It is the minutes a senior human can spend on what only they can sign.Same constraint in regulated services. The attention worth spending is the partner reading the regulator memo, the client letter, and the filing before it goes out. The binding number is not compute or seats. It is the minutes a senior human can spend on what only they can sign.

@steipete Agents talking to each other cleanly is the part most teams skip. We run group chats where one agent triages, one drafts, and one runs QC. The visible replies pattern is the only thing that keeps the partner reading the channel without translation.

@ryancarson Monthly drill is the part most teams put on the wishlist forever. We schedule the restore on the calendar, run it on a dummy matter, and the partner who reads the timing report is the one who signs off the runbook. The drill is the only honest test the harness ever takes.

Same shape in legal services. The work that carries criminal liability is the work nobody is vibe coding. Firms keep paying for the compliance memo and the regulatory filing because the penalty is borne by a human with a practising certificate. AI changes how the work gets done. It does not change who signs the letter.Same shape in legal services. The work that carries criminal liability is the work nobody vibe codes. Firms still pay for the compliance memo because the penalty is borne by a human with a practising certificate. AI changes how the work is done. It does not change who signs the letter.Same shape in legal services. The work that carries criminal liability is the work nobody vibe codes. Firms still pay for the compliance memo because the penalty lands on a human with a practising certificate. AI changes how the work gets done, not who signs the letter.

@kxie06 Private models on the customer's infrastructure is the part most platform plays still skip. We run legal agents that route between API and local based on matter sensitivity. The teams shipping in regulated work are the ones whose stack a regulator can audit on a Tuesday morning.

The hiring binge plays out cleanly in legal. The firms that scaled associates through twenty twenty two are sitting on operating models built for a market that no longer exists. AI did not cause that. It just made the model show up on the P and L faster than the partners wanted to read.

@openclaw The Ollama and Hermes carry over is the part most platform releases skip. We run legal agents that switch between API and local depending on matter sensitivity. The harness does not get rebuilt. Local models eat well is right.

@levie The last mile is where every regulated services play wins or loses. We run agents that triage matters, draft positions, and prep filings. The partner still owns the regulatory call, the client tone, and the sign off. The firms automating the judgment layer keep blowing up.

@HarryStebbings The same compounding works in regulated services. When a partner spends an hour a day on side matters, the loss does not show up in the quarter. It shows up two years later in the operating model that did not get sharpened. The discipline that pays reads boring on paper.

@vkhosla Same shape in regulated services. The danger is not the burn. It is a year of compliance overhead baked into the operating model. The firms running agents under associates can change cost structure inside a quarter. The legacy firms cannot brake without losing clients first.

@ycombinator @t_blom The skills file is the part most teams mistake for documentation. We run legal agents off a markdown directory per matter, a daily log per agent, and a rolling index the partner reads without translation. The brain is not the wiki. It is what the agent reads before every task.

Same lesson in legal services. The founders worried about Anthropic eating their legal startup are arguing about the wrong failure mode. The firms that fail will fail because they never changed the operating model, not because a lab ran them over.

Paul Graham @paulg

2 months ago

Worrying that your startup will be eaten by the model companies is like worrying that your life will be constrained after you become a movie star. You're far more likely simply to fail.

248 347 6K 312K 959

The DR exercise is the part nobody schedules until the first time they need it. We run a quarterly drill where one of the agents is told to act on a dummy matter that mirrors a live one, and we time how fast we restore. The drill is the only honest way to know the runbook still works.

The credential boundary is the part most teams discover after a destructive delete, not before. We run legal agents with read replica access and a hard wall to anything that mutates client data. The harness assumes the agent will go wrong. The audit trail is the only place you find out it did.

The OTEL bit is the part that quietly changes how we run agents in regulated work. We needed to know which legal agent did what, when, and on whose data, with the audit trail that compliance can read without translation. Less mystery is the prerequisite for shipping in matters that matter.

@ArtificialLawya @AnthropicAI The asymmetry is the part most coverage will skim past. Stronger model representing you negotiates a better outcome. Same dynamic in litigation. The firm running cheap models on matters walks away with less on the table than the firm that paid for the heavy one.