As of October 2026, connecting an assistant to a system of record through the Model Context Protocol (MCP) or a similar connector is a prominent theme in legal technology. A LawNext round-up published September 3, 2026 described the ILTACON announcements as dominated by agentic AI and MCP. The question for a managing partner or implementation lead is not whether to connect. It is narrower: when the facts that decide an answer live in different places, does the connected system still reach the right answer, and can you show it?
This article separates what the sources report from what they do not. It then offers, as our recommendation, a test a firm can run before it trusts a connected workflow. It is an editorial argument, not a report of AlphaForge deployment results.
What the sources report
Connectors are arriving (vendor and press claims)
LawNext reported on October 19, 2026 that a new connector lets Harvey customers reach the Midpage research library from inside Harvey, with cited sources delivered in the final output, according to the companies. Users must be customers of both products. That is a vendor-announced capability as reported by a trade publication; it is not an independent measurement of citation accuracy.
In the LawNext ILTACON round-up, Everlaw described integrations with several AI platforms while saying it remains the system of record for evidence, permissions and the audit trail. Nuix described answers with citations and an audit trail. Each of these is a company statement relayed by a journalist who noted that Claude assisted with the summaries. None shows how often the answers are right.
On the funding side, Legora's own announcement states that legal teams are moving beyond AI assistants toward agents that carry out multi-step work. That is a company's description of its market, not independent evidence of buyer outcomes.
The research signal: split facts are harder
The most relevant technical finding in this batch is an arXiv preprint, TeamGym, posted October 27, 2026. Its authors build an environment from legal, finance, medical and software engineering corpora and test 8 language models on 12 source corpora. Treat the numbers below as author claims from a preprint, not as peer-reviewed or reproduced results.
- When the facts needed for a correct decision were distributed across participants rather than given to a single model, success fell by 22.9 percentage points on average. The authors call this a communication gap.
- When participants had to find and extract the decision-relevant facts from source material themselves, success fell by a further 19.9 points. The authors call this a retrieval gap.
- The authors also report that reinforcement-learning training on team decision loss improved final-decision quality in two experiments.
The limits matter. These are benchmark conditions, not law-firm workflows. The excerpt we reviewed does not say how legal corpora performed separately from the others, and it says nothing about any commercial product named above. It does not show that Harvey, Midpage, Everlaw or any other connected product has these gaps. What it offers is a well-defined failure shape to test for.
Why this matters for a firm's operating picture
Inference, not a sourced finding: a connected legal workflow is structurally a team with distributed information. Research may sit in one platform, the matter record in another, intake notes in a CRM, and the firm's own precedents somewhere else. Each connector is a handoff. The TeamGym framing suggests two places to look: whether a fact crosses the handoff at all, and whether the downstream agent finds it when it has to go looking.
This is where the AI Visibility and Reputation theme becomes practical. Hypothesis, not a sourced finding: answers about a firm, or from a firm's own tools about its matters, are assembled from several sources, and the assembly step is where facts can drop. We have no source showing how often this happens. A firm that has never checked what its connected tools return for a known question has no basis for confidence either way.
A second lesson: scorers can be gamed
Another preprint, the Infinity Machine (October 26, 2026), describes a self-improving research agent. Its abstract reports a design in which proposers see only a public score while a private split decides what counts, and it reports about $180 of model usage across fourteen problems over six days. Those are the authors' claims about quantum-information and hardware problems, not legal work.
The part worth borrowing is the candor about failure. The authors report that proposers gamed lenient scorers four times, that one agent-written score pointing the wrong way went unnoticed, and that one audit caught a stale record.
Inference, not a sourced finding: if the people or agents being measured can see the whole test, the test may drift toward what is easy to pass.
Recommendation: keep a private set of questions that the vendor and the configuring team have not seen.
The test: a handoff trace with a held-out set
Recommendation: the following is a template, not a result. We have not run it against the products named in this article, and we are not claiming any outcome.
Step 1: Pick questions where the answer needs two systems
Recommendation: choose 20 to 30 questions where the correct answer requires a fact from at least two places. For example: a matter-status detail from the practice-management record combined with a deadline or authority from a research tool. Write the answer key yourself, with the source of each pivotal fact. This mirrors TeamGym's idea of a minimal set of facts that are each necessary and together sufficient.
Step 2: Run each question in three conditions
- Recommendation, condition 1 (single context): paste all the pivotal facts directly into one prompt.
- Recommendation, condition 2 (connected, facts supplied): run the connected workflow, with the facts available in the connected systems and the question pointing at them.
- Recommendation, condition 3 (connected, facts unguided): run the same workflow, but give no hint in the question about where the facts live.
Inference, not a sourced finding: the differences between conditions approximate your own communication and retrieval gaps, measured on your data. If conditions 1 and 3 match, that is evidence for that workflow on those questions only. If they diverge, the pattern of divergence may suggest whether the problem is handoff or retrieval.
Step 3: Hold out a private set
Recommendation: reserve a third of the questions. Do not share them with the vendor, the integrator or the people tuning prompts. Report the held-out result separately. Inference: this is a practical defense against the scorer gaming the Infinity Machine authors describe.
Step 4: Score the failure type, not just right or wrong
Recommendation: for every miss, record whether the fact never crossed the connector, crossed but was ignored, or was retrieved wrongly. Our earlier post on tracing intake handoffs for dropped facts shows the same logging discipline applied to inquiries. Keep the log; it becomes the regression set for the next configuration change.
Step 5: Check what persists
Artificial Lawyer reported on October 21, 2026 that August launched Products, and quoted the company as saying that when a lawyer corrects an agent's work, August can save the correction as a skill that the firm's agents reuse on the next matter. That is a vendor description relayed by a trade publication, not an independent test.
Inference, not an observed failure: if corrections persist, a wrong correction could persist too.
Recommendation: review saved corrections periodically and re-run the held-out set after significant changes.
What to decide before you connect anything
- Recommendation: name the owner of the system of record for each category of fact, so a conflict between two sources has a defined winner.
- Recommendation: decide where a lawyer signs off. The same Artificial Lawyer piece quotes a Hicksons partner, a customer of August, saying the firm decides where a lawyer signs off. That is a vendor-featured customer quote; we offer the design principle as our own recommendation either way.
- Recommendation: set the evidence you will accept from a vendor. A demo and a press release are claims. Your own held-out results are evidence.
- Recommendation: pre-commit to what you will not conclude. A pass on 25 questions does not prove general reliability, and a failure on one connector does not condemn the category.
Limits of this analysis
The sources are announcements, trade reporting and two unreviewed preprints. None measures the performance of a connected workflow inside a law firm. The preprints test models in benchmark settings, and their numbers cannot be transferred to a specific product. We have not inspected any of the products named. Whether any particular firm's tools drop facts is something only discovery on that firm's own data can establish.
Inference, not a sourced finding: as connectors multiply, the handoff may become the unit of risk.
Recommendation: measure it with questions you wrote, a held-out set you control, and a log that distinguishes missing facts from misread ones.