Lawrence Experience · Cycle 11 · 6 – 19 August 2026

Sprint 11

Lawrence 2: Parity, Cutover-Ready

Final
Lawhive

Sprint goals

Finish M3 (Lawrence 2 fully on the TypeScript runtime), plus identity, archiving, and forms polish.

32 issues landed · git-verifiedcounted from trunk merges, not from Linear

Landed points by sprint

~76
S7
102
S8
93
S9
65
S10
~72
S11 · final

32 issues landed, against 44 in Sprint 10.

7
issues in Prod
25
in Dev
+6
one merge away · open PRs in review
160 120 80 40 0 Aug 6 Aug 19 Aug 10 backlog import git-verified: 32 issues, ~72 pts landed on trunk Linear counts only prod deploys as completed committed scope · 168 started · 82 (49%) completed · 29 (17%)

The 10 August jump from 40 to 110 points is backlog housekeeping: roughly 14 old tickets were imported into the cycle, so it is not real scope growth.

Lawhive

Release

Lawrence 2 is live for everyone

100% of v3 users

Rollout · Lawrence 2 chat on v3

Every v3 user 100%

Rolled out to every v3 user on 12 August.

  • Every lawyer on the platform now works with Lawrence 2 chat: grounded in the matter, with forms, drafting and citations.

The release runs on the Python runtime. Moving it onto TypeScript is the next step, and a separate one.

Lawhive

Goal

Lawrence 2: full parity on the TypeScript runtime

Done · cutover pending
✓ Read and search tools landed last sprint ✓ Write verbs create, edit, ask the user ✓ Skills and citations real skills, citable sources ✓ Prompt parity context, params, step budget ✓ Routing flag runtime chosen per user
  • All three write tools landed: ask_user_question as a client-fulfilled primitive, plus create with stat, plus edit Editing now reaches email drafts, custom field definitions and matter template application.
  • Real skills replaced the demo ones: legal research with its market variants, file organisation, form filling, precedent filling, drafting
  • Chat citations ported: citation models, per-tool citable sources, and markers parsed into structured references
  • Prompt parity: template user message, the caseContext shape, Langfuse model params with prompt caching, a dynamic step budget, error spans for invalid tool inputs

In review: create-tool validation (LEX-890/891). Next: QA with Sedona and Harry, then release, then cleanup of the old runtime.

Lawhive

Agent Platform

One flag away from a single runtime

legal-os client
lawrence-api
Feature flag · LEX-887
routes chat per user
Python runtime
agents repo. Serves production today.
TypeScript agent platform
agent-worker, durable journaled runs. Parity complete.

Flip per user, roll back instantly.No data migration, no big-bang release.

What this unlocks

One runtime to maintain instead of two Durable, journaled runs with full tracing Shared TypeScript components across app and admin The old Python chat path can be retired
Lawhive

Demo

Global agents: the agent bar and tabs

Lawhive

Goal

Every action now answers: who did this?

Done for this phase
  • Open any task, note, draft or document and it can say who created it. That is recorded, not inferred.
  • It records whether the work was done by hand or with Lawrence: the person and the agent are captured together.
  • Each agent action links back to its run: what was asked, what the agent did, what it produced.

Under the hood: an agent identity type, one seeded identity per agent, and author plus executing-agent columns across the matter, email and content services.

Parked deliberately until background-agent configuration resumes, so not next sprint.

Lawhive

Goal

Quality of life: archive threads shipped

Done
  • Archive from the panel header. Archived threads get their own group in All chats, and you select one to restore it.
  • Sends to an archived thread are rejected, the agent can see that a thread is archived, and archiving is disabled mid-stream.
  • Stop actually stops: STOPPING and STOPPED statuses, a stop route, the worker ends the turn, and the tool calls stay visible.

Not started, and carried: the delete pattern, the canvas pop-up, and making Lawrence optional.

Lawhive

Around the edges

Also this sprint

Release autopilot
WIP · coming soon Will propose and cut production releases from Slack, approved with a reaction.
Admin Lawrence tab
A thread picker and full message rendering, through the shared lawrence-components.
VFS reads hardened
The read routes now require signed requests.
Prompt suggestions
Grounded in a digest of recent matter activity.
Forms polish wall
Preview states, licence gating, virtualized suggestions, review-state fixes.
Cross-squad credit · Libaan Hassan, LegalOS LEX-863: fixed the stop route 500.
Lawhive

What's next

Next sprint

Pending sprint planning
Agent Platform
QA, then the cutover to the TypeScript runtime, then cleanup of the old runtime. M2, agents beyond matters, starts building from the published plan.
Forms
Agree the direction for the first-principles redesign.
Team
Adolfo, Dave and Pratik move to table-stakes work. Adolfo on the migration engine; Dave and Pratik supporting BOS on legal cashiering and invoicing.

Carried over

Delete pattern Canvas pop-up Lawrence optional LEX-890 / 891 Identity tail · parked
Lawhive

Thank you

Full report on Notion → LEX Sprint 11 Report Sprint 11 goals → Notion
Lawhive