Multi-agent version control with reactive feedback

alansaber1 pts0 comments

Multi-agent version control with reactive feedback | Lexifina<br>Skip to main contentFeaturesSecurityPricingBlogChangelogSign in

Managing agent collisions when working on a single document.

There will be two kinds of multi-agent workflows in use: orchestrated swarms, where goals and context are cleanly planned and delineated for each agent; and ad-hoc agent use, where users accumulate multiple agents working across similar areas with potential for collisions. Most real use follows the second case.

Lexifina active-agent coordination architecture<br>A proposal reaches a target-scoped DB staging gate. Disjoint work commits immediately. Relevant peer work causes that proposal attempt to fail without writing, returns an exact peer and intent packet to the same model run, and durably records a completion challenge. The agent can retry as a verified complement, a related edit, an explicit alternative, or a provably disjoint proposal. The successful transaction atomically advances target heads, inserts the proposal, and clears the challenge. Frontend wakes remain asynchronous and do not enter active model context.

ACTIVE AGENT<br>STAGING BOUNDARY<br>DURABLE STATE<br>OBSERVERS + REVIEW

Candidate proposal<br>pinned document version, revision + hash<br>exact target and declared basis sections<br>typed intent, objectives + constraints

NO LEASE WHILE THINKING

Target-scoped staging gate<br>derive target, section, basis + formatting fences<br>read current target heads + relevant peers<br>transaction rechecks document, peers + revisions<br>compare-and-swap only intersecting target heads

NO DOCUMENT-WIDE AGENT COUNTER

Atomic stage<br>advance the relevant target heads<br>insert immutable proposal + exact receipt<br>clear the resolved run challenge

DISJOINT TARGETS STAY PARALLEL

Post-commit wake<br>coalesce board revision + proposal IDs<br>agent.proposal.changed

Lawyer's shared review<br>separately attributable proposals<br>explicit complements and alternatives<br>approved set enters document version control

THE LAWYER MAKES THE LEGAL DECISION

NO ACTIVE PEER

REPLAN<br>Same run adapts

verified complement

explicit human-choice alternative

provably disjoint target

COLLISION PATH<br>Refuse this attempt<br>write no proposal row; return exact peer evidence:<br>before / after text + source session<br>intent, required objectives + constraints<br>current revisions + intent / effect digests<br>VALID RESPONSE ENTERS THE SAME MODEL RUN

DURABLE DUTY<br>Run challenge<br>exact packet + digest stored on the run<br>survives worker ownership changes<br>natural stop triggers up to 3 exact nudges

ACTIVE PEER

EXACT PACKET

RETRY

SYSTEM INVARIANTS

READS STAY PARALLEL<br>ONLY INTERSECTING TARGETS SERIALIZE<br>NOTIFICATIONS ARE POST-COMMIT<br>ONLY APPLY WRITES DOCUMENT HISTORY

ACTIVE AGENT<br>COORDINATION BOUNDARY<br>AUTHORITATIVE BOARD<br>WAKES + REVIEW

Private run context<br>same pinned canonical document<br>own instructions, reasoning and draft intent

Choose a drafting target<br>selected naturally from the whole document<br>exact target, basis sections and objectives

Same run adapts<br>move to provably disjoint work<br>submit a verified complement<br>retain an explicit human choice

Target-scoped staging gate<br>propose_edit enters this gate automatically<br>read exact target heads + relevant peer frontier

No relevant peer proposal<br>stage immediately in the existing transaction<br>board CAS handles concurrent persistence<br>FAST PATH · NO EXTRA MODEL INVOCATION

Relevant overlap found<br>reject only this attempt; write no proposal row<br>return exact text, intent, objectives and digests<br>TOOL ERROR ENTERS THE SAME MODEL RUN

retry after adaptation

DB target heads + board<br>per-target course CAS

Target heads<br>section, basis, formatting and parent fences

Proposal rows<br>immutable lineage + exact stage receipt

Run challenge<br>cleared in the successful stage transaction<br>DISJOINT TARGETS DO NOT CONTEND.

Unresolved collision<br>peer evidence saved with the run<br>recovery survives worker changes<br>NO FALSE SUCCESS

Post-commit observer wake<br>coalesce board watermark + proposal IDs<br>agent.proposal.changed refreshes the UI only<br>NEVER DELAYS OR REWRITES A SUCCESSFUL STAGE

Lawyer's shared staging review<br>all proposals remain separately attributable<br>complements and alternatives remain explicit<br>accepted set commits through document VCS<br>HUMAN RETAINS THE FINAL LEGAL DECISION

SHIPPED · TRANSACTION-GATED ACTIVE COORDINATION<br>No polling, no mid-token interruption, no public board-read tool: relevant peer state is returned only when it can change a staging decision.

SAME ACTIVE RUN<br>TOOL BOUNDARY<br>AUTHORITATIVE COORDINATION GATE<br>DURABLE STATE<br>OBSERVERS + LAWYER

Candidate proposal<br>pinned version + revision + content hash<br>exact target and declared basis sections<br>typed intent, objectives and must-preserve rules

NO BOARD CURSOR OR BACKGROUND READ

Proposal tool call<br>propose_edit / propose_edits<br>candidate reaches one safe boundary

Target-scoped staging gate<br>derive fences from target, section UUID / ID,<br>basis sections, formatting scope and parent IDs<br>read target heads + relevant active...

target proposal agent exact document active

Related Articles