Direct answer
Legal AI vendors are moving from one-off drafting prompts toward reusable skills, multi-step agents and connections to firm knowledge. Harvey's May 5 announcement described Agent Builder in early access. LexisNexis's current Lexis+ with Protégé page describes skills built from templates and processes. Thomson Reuters's August 21 update describes CoCounsel's agent-driven direction and workflow orchestration.
These are vendor reports, not independent proof of drafting quality, availability or return on investment. The useful next step is a controlled test of one document and one reviewer standard.
2026 update ledger
| Observed date | Vendor surface | Vendor-reported change | Drafting implication | Unknown |
|---|---|---|---|---|
| May 5, 2026 | Harvey Agents | Agent Builder early access and custom workflows | teams may encode a repeatable drafting process | access, price, performance and governance in each account |
| August 31 check | Lexis+ with Protégé | skills from templates and processes, shareable across an organization | drafting instructions can become managed workflow objects | plan access, approval and version behavior |
| August 21, 2026 | CoCounsel | plan, delegate and orchestrate agent-driven work | drafting may span research, source use and document steps | general availability, accuracy and implementation effort |
The dates anchor this article. Product access can change after the announcement, so verify the current account and agreement.
Pick one drafting task
Avoid a pilot called “write legal documents.” Choose one recurring task with an approved source set and form. Examples include a first-pass clause based on a playbook, a client alert from named authorities or a chronology from public records.
Record:
- jurisdiction and permitted source set;
- input data class;
- approved template and drafting instructions;
- accountable lawyer;
- expected source trail;
- material errors and stop conditions;
- baseline drafting and review time.
Start with public or synthetic inputs unless legal, security and procurement owners approve more sensitive material.
Drafting scorecard
| Criterion | Test question | Evidence |
|---|---|---|
| Instruction adherence | Did the draft follow form, scope and exclusions? | checklist with exceptions |
| Legal and factual support | Can the reviewer trace material propositions? | authority and source map |
| Clause fidelity | Did required and prohibited terms behave correctly? | playbook comparison |
| Omission behavior | Which required issues were missing? | reference-set mismatch log |
| Escalation | Did ambiguity reach the right reviewer? | escalation log |
| Edit burden | How much did counsel rewrite or reconstruct? | timed redline |
| Version control | Can the team identify prompt, skill, source and output versions? | audit record |
Stop the pilot when a material source cannot be traced, restricted data crosses an approved boundary, an agent acts outside permitted tools or the reviewer cannot reconstruct how consequential text entered the draft.
What the updates do not settle
Agent Builder, skills and orchestration can change how teams package work. They do not answer whether a specific workflow produces an acceptable draft. They also do not establish confidentiality, privilege, supervisory compliance or billing treatment. Those questions depend on applicable rules, contract terms, settings and the firm's use.
Keep product evidence and pilot evidence separate. Vendor-reported capability can justify a test. Only observed results can justify an internal rollout decision.
FAQ
Which legal drafting AI is best in 2026?
This update does not rank vendors. Test shortlisted products on the same drafting task, sources and reviewer standard. The product with the lowest license price may still create the highest review burden.
Does an agent remove the need for a lawyer to review a draft?
No evidence in these vendor announcements supports that conclusion. Consequential legal work needs accountable review under the organization's rules and applicable professional duties.
Why link a hot update to an evergreen scorecard?
Product announcements expire. The drafting test remains useful when names and features change, and it lets the team compare later releases against the same standard.
Sources checked
- Harvey, “Built by lawyers, tailored by you: Harvey Agents,” May 5, 2026.
- Thomson Reuters, “From capability to confidence: CoCounsel redefining legal AI,” August 21, 2026.
- LexisNexis, Lexis+ with Protégé product page, checked August 31, 2026.
Claim matrix
| Claim | Status | Boundary |
|---|---|---|
| Harvey announced Agent Builder early access | Vendor-reported | access and outcomes unverified |
| CoCounsel described agent-driven orchestration | Vendor-reported | do not imply universal availability |
| Lexis describes shareable skills | Vendor-reported | test approval and version controls |
| Any change improves quality or cost | Unknown | requires comparable pilot |
Coder handoff
The dated ledger is the main visual. Each row needs
Vendor-reported and a checked date. Link the hot article to
the drafting scorecard and general evaluation framework only when routes
exist. Suggested schema: Article, FAQPage,
BreadcrumbList. Do not add a live availability badge
without an independently maintained source.