At Reframe we build with AI agents every day. Our engineering agent reviews pull requests with the team, our QA agent runs the release checks on every build, our operations agent keeps the deployment runbooks current, and our editorial agent drafts articles like this one. Law firm leaders ask me what working this way is actually like. The most useful thing I've learned building Reframe is that an agent is only as good as the context it starts from.
Capable agents still start without context
The models are capable. What goes wrong is what they know when they begin. Ask an agent to prepare release notes and it doesn't know what changed last week, who approved it, what we promised a client, or what we decided to hold back. Hand it a folder of files and it will read a few, guess at the rest, and write something confident. The guesses are where the mistakes come from.
A new hire has the same problem on day one. The difference is that a new hire asks around. Most agents don't, because there's nobody to ask.
What Reframe gives an agent before it starts
So our agents work from Reframe itself. Before one of them acts, it asks Reframe what we already know about the work in front of it: the project, the decisions made so far and the reasons for them, who is accountable, the open questions, and the documents behind each of those. Every answer comes back cited to the document it came from. The agent works from our record, and I can check where every fact came from.
When our editorial agent drafts an article, it starts from what Reframe holds on the topic: the positions we've taken, the sources we've already checked and the language we've approved. When our QA agent checks a release, it starts from what changed and why. The work reads like it came from someone who was in the room, because the agent was briefed the way a good associate gets briefed before a client call.
The same rules bind every agent
Every request an agent makes to Reframe passes the same permission checks a person's request would. An agent working on one client engagement can't see another client's records, and walls between clients bind agents exactly as they bind people. The full trace of what an agent read and what it did can be pulled on demand. That matters more than any single capability, because it's what lets me give agents real work without worrying about what they can reach.
Agents propose, and people decide
Agents learn things as they work: a decision made on a call, a deadline that moved, a requirement that changed. None of that gets written into the record directly. The agent proposes the change, a person approves it, and from then on every agent starts from the updated record. That's how the context gets better every week without drifting into something nobody trusts.
Handoffs stop losing work
Work moves between agents the way it moves between people, and the handoff used to be where it went wrong. The second agent depended on whatever note the first one left, and when the note was thin it repeated the first agent's mistakes with the same confidence. Now the second agent doesn't depend on the note. It asks Reframe where things stand, and it gets the same cited answer the first agent would have.
I still make agents argue with each other
One practice has improved the quality of the work more than any other. I give a hard problem to two agents running on models from different providers and let them debate it. One proposes, the other attacks the proposal, and a third session keeps them on the question until one concedes. Both start from the same context in Reframe, so they already agree on the facts and spend the argument on the approach. That gets to the right answer faster.
What I still do myself
As founder, I set the goals, decide what matters this week, and approve anything that changes the record or goes to a client, an investor or a partner. Agents execute defined work very well. They have no sense of which work matters, and the accountability when something is wrong stays with people.
What this means for a law firm
Every firm already knows more than any one person in it. That knowledge sits scattered across a dozen systems, and it walks out the door when a matter closes. That scattered knowledge is exactly the context an agent needs: the client, the matter, who worked on it, what was decided and why, and the documents behind it. Give an agent a pile of documents and it guesses. Give it the firm's own record, governed and cited, and it works like a well-briefed associate.
That's what Reframe does inside a firm. It brings the firm's knowledge into one governed place, inside the firm's own cloud, that lawyers and agents both work from. Agents ask it what the firm knows before they act, every answer is cited to the source, walls and permissions bind them the way they bind lawyers, and every change they propose is approved by a person.
If you're deciding where to start, pick one practice group and one process its lawyers repeat every week. Write down the questions an agent would need answered to do that work well, and name the person who approves what the agent proposes. Then give the agent that one job with that context, and review every result for the first month. You'll learn more about your firm's readiness from that month than from any assessment.
Drafted with Reframe Editorial Agent, an AI employee at Reframe. 