Agent Branches
Geometric blue agents carry folders along branching commit lines into an orange merge

Day 13: I Recorded the Demo and Submitted the Entry

I submitted my entry after changing the pitch: not a better Git, but a coordination layer that warns agents before they overlap. After the recording, I asked agents to add planned-file warnings and a watcher, reviewed but not yet on Cloudflare.

Alexey GrigorevWritten with Claude Opus15 October 2026EVIDENCE UP TO 2026-10-15 00:09 UTC

Read the daily report

The hypotheses

2 still being tested. 18 parked. 4 places open. None selected. Every idea is provisional. Read what all 20 ideas would do

PROVISIONALLive Integration RadarPrimary recommendation withdrawnProblem: Parallel changes can pass their own tests and still fail when combined.Idea: Can unfinished agent work reveal a composition problem before the tasks are complete?PROVISIONALStorage-aware workspacesParked as competition candidate; final shortlist unsignedProblem: Separate agent workspaces can fill the disk with repeated dependencies and builds.Idea: Keep writable tasks isolated without multiplying dependencies and build footprints.PROVISIONALBehavior-based fork tournamentParked provisionally; selection advantage not demonstratedProblem: Comparing several agents' attempts takes time, and a plausible answer can still be wrong.Idea: Choose between parallel attempts by what they do, with a bounded and independent review.PROVISIONALChange-story reviewInternal use; comparative benefit unprovenProblem: Agent-generated changes increase review work; long summaries can hide mistakes or mislead reviewers.Idea: A short account of behavioral changes, tied to the evidence a reviewer actually needs.PROVISIONALDurable task handoffParked provisionally; ordinary recovery baseline sufficedProblem: When an agent restarts or hands off a task, its intent, unfinished work and correct code version can get lost.Idea: Resume the right task on the right code state without reconstructing an entire conversation.PROVISIONALLease-bound claims enforced at landingParked; reopen only if the expired-lease race test passesProblem: Stop two agents from silently doing overlapping work.Idea: an agent claims the files it's working on, and the system refuses to merge overlapping changes from another agent.PROVISIONALMerged-state gate + bounded intent reapplicationParked; reopen only after the merged-state fixture testProblem: A change made against an older code version can break newer work when it is merged.Idea: the system tests changes together before it accepts them.PROVISIONALSemantic contract sentinelParked; compared against running the full test suite on the merged stateProblem: Changes can merge without text conflicts but break behavior that other code depends on.Idea: catches changes that merge cleanly but quietly break something another part of the code relies on.PROVISIONALMaintainer inbound quarantineParked; reopen test recorded, not selectedProblem: Maintainers spend time reviewing AI contributions that lack a working fix or evidence.Idea: contributions from AI wait in a holding area until they show they're worth a human's time.PROVISIONALIndependent reviewer panel on pushNot selected; parked without a testProblem: Bugs in agent-written code can reach human review without an independent check.Idea: AI reviewers unrelated to the agent that wrote a change check it before a human looks.PROVISIONALExact-SHA verification receiptsParked; the narrowed publication-receipt idea waits on a kill testProblem: A passing test report may refer to an older change or a different combination of commits.Idea: a trusted record of exactly which version someone tested, under which rules.PROVISIONALMerge decision ledgerNot selected; parked without a testProblem: Git history can show what changed without explaining why or which alternatives were rejected.Idea: for any piece of code, it records why it was written that way and which alternatives were rejected.PROVISIONALQuarantine forks + capability tokensParked; Task Passports refinement waits on first-hand demandProblem: An agent with broad Git access can overwrite shared history or affect another agent's work.Idea: agents work in separate copies with limited permissions, so they can't delete or overwrite the main history.PROVISIONALSession footprint undo across forksParked; reopen test recorded, not selectedProblem: One agent's changes can spread across several branches, making them hard to find and undo safely.Idea: you see everything one agent did, everywhere, and reverse it in one step.PROVISIONALPreview-per-agent runtime isolationParked; folded into the collision radar as runtime verificationProblem: Agents running separate changes can still collide over ports, databases and shared runtime data.Idea: each change gets its own running preview and data, so agents don't fight over the same ports or database.PROVISIONALPush-triggered fast verification laneNot selected; parked without a testProblem: Slow or expensive test runs delay feedback, and agents can miss failed checks.Idea: after every change, the system runs only the tests the change could affect and reports back within seconds.PROVISIONALFork-is-the-task boardNot selected; parked without a testProblem: People must keep task-board status in sync with the code agents are actually working on.Idea: there's no separate task board.PROVISIONALMulti-repo change sets with recoveryParked; investigate-first, must beat the existing deploy-check toolsProblem: Changes across related repositories can leave incompatible versions if integration fails partway.Idea: one change rolls out across several related codebases in the right order, with a way to recover if it fails partway.PROVISIONALMaintenance swarm with batch landingParked; partly folded into the collision radarProblem: Many small maintenance changes can create a backlog of test runs and merges.Idea: dozens of agents each take one chore, such as an upgrade or a cleanup, and the system merges the compatible results together.PROVISIONALEarned-autonomy policy per agent configNot selected; parked without a testProblem: Reviewing every small agent change takes time, but letting all changes merge automatically risks defects.Idea: setups with a good track record can merge low-risk changes on their own, and everything else goes to a human.
OPENOpen place — nothing selectedReopened · no candidate approvedThe private test site idea was folded into the collision radar. Nothing has earned the sixth place yet.

Most of the worktree pile isn’t Git. It’s dependencies.

The pain is real and measured. Whether anyone would adopt a product for it is unknown. Package managers that already keep one shared copy of dependencies for many folders may solve it without a new product.

Read the storage-aware workspaces idea
472linked worktrees
25repositories
111.7GiB on disk
62.1%dependencies and build output
Dependencies and build output: 69.4 GiB (62.1%)Source code and everything else: about 42 GiB
Whether this needs a new product: unknown

Research library

All sources

Everything links to a file in the public repo or its archive tag.

Shortlist draft (unsigned)research/shortlist-6.mdWorktree disk on the real hostresearch/claude/u7-real-worktree-measurement.mdConsensus record — pendingresearch/consensus.mdAll 20 approachesresearch/approaches-20.md