session start · “what should I do next?”
startMigrate session storage · ready · blocks 3
heldWire invoice emails · dependency open
One trustworthy answer to what happens next — even while you and your agents change the work. Sift Q refuses competing claims, blocks unready tasks, and protects work already in flight.
Race the write ↓Your coding tools run the agents. Sift Q keeps their plan true. · Request early access
Next action · ready work only
.92Migrate session storageready · blocks 3
.87Fix mobile token refreshcaptured by you
—Wire invoice emailsheld · waiting on #138
—Polish billing settingsclaimed · agent 2
Task #138 · one live claim
held by agent · claude
heartbeat renewed · lease active
if the heartbeat stops, the task returns to the pool — never falsely marked complete
Two writers · task #138 · v7commit either first
youclaim “Migrate session storage”
agentclaim “Migrate session storage”
one claim lands; the stale claim gets a clear no
Direction changed · previewed first
re-ranked “stabilize auth before billing polish” — in-flight migration protected, uncertain work left pending
Your coding tools run the agents. Sift Q keeps one coherent answer to “what happens next?” while humans and agents keep changing the plan.
One collision, handled
Two writers reach for the same task. Only one starts.
The circle is you. The squares are agents. When both claim the same work, Sift Q grants the first claim and gives the stale one a clear no.
Try the collisionYou and an agent both claim “Migrate storage.”
seeded local simulation · no backendWhat you’re seeingOne task. Two possible writers.The first claim lands. A competing claim is refused before duplicate work starts.
youagent
The Value, Before the Mechanism
What changes on Monday
you · “we need a regression test for mobile token refresh”
captured#144 · auth · effort low · waits on #138
current#138 stays in progress · no tracker detour
Capture the tangent. Keep your flow.
directive · “stabilize auth before billing polish”
preview4 promoted · 1 protected · 3 pending
applyconfirmed with receipt · undo available
Change course without rug-pulls
#151 · lease holder silent
last heartbeat · expired
last heartbeat · expired
returnedtask is claimable · history preserved
outcomestill open · completion never invented
Recover work after an agent disappears
One Seeded Project, Start to Finish
A day with Sift Q
08:42 · Ask what’s next
The agent gets “Migrate session storage”: the first ready task, not the loudest blocked card.
11:05 · Claim before building
A second agent tries the same migration. Its claim is refused before a second branch exists.
14:10 · Change direction safely
Auth moves up. The active migration stays protected, the change is attributed, and undo remains available.
The Mechanism, in the Open
Not a thin wrapper over MCP
youwrite(#121, base: v7) ✓ → v8
agentwrite(#121, base: v7) ✗ stale — v8 exists
rejected, surfaced, attributed. never clobbered.
No double-grant
#138 · lease: agent-claude
heartbeat ·· 4s ago
heartbeat ·· 4s ago
t+60sno heartbeat → lease expires
#138 · back in queue
a dead agent can’t hold your work hostage
a dead agent can’t hold your work hostage
No ghost ownership
.92ship auth flow
.87cut billing scope
heldwire invoice emails — dep #142 not done
agents only see work that’s actually ready
No plausible-looking blocked work
#144 · rotate service credential
notes: value iscredential-value inruntime config
notes: value iscredential-value inruntime config
boundarywhat agents can read is a boundary, not a hope
No accidental context spill
The Directive, Explained
Change your mind without thrash
↑1Ship auth flowpromoted
=Migrate session storagein flight · protected
↓4Polish settings pageagent notified
Priorities change mid-run. Preview the movement first, protect work already in flight, expose uncertain classifications, then apply the new direction with a receipt and a route back.
- You state the new directionplain language, one line→
- Sift Q previews the movebefore, after, rationale, uncertainty→
- In-flight work is protectedno mid-task rug-pulls→
- You confirm or undodigest-bound apply, durable receipt→
The Boundary
Local is free. Shared work is the upgrade.
Local forgeShared hub
Free/local forge
Tasks, ranking, dependencies, and lifecycle stay available locally. Shared coordination is the paid upgrade. Exact hosted onboarding and pricing will be confirmed before this page launches.
Request design-partner access →What the shared layer adds
- A shared task pool across humans and agents
- Atomic claims and typed contention
- Current focus, blocker edges, and handoffs
- Leases, heartbeats, and abandoned-work recovery
- Attribution and lifecycle history
- Local queues and ranking signals stay local
One startable next action
Leases, not assignments
Useful at one agent
Tangents captured without a detour
Conflicts surfaced, not swallowed
In-flight work protected
Local judgment stays local
Re-rankable mid-run
FAQ
Your questions, answered
Is this another orchestrator?
No. Your coding tools run agents. Sift Q tracks and arbitrates the work they and your team change together: what can start, who holds it, what happens under contention, and why the order moved.
I run one agent. Is this for me?
Yes. Solo value is immediate: capture off-path work without leaving the current task, keep blocked work out of the answer, and start each session with one genuinely startable action.
What happens when we both edit the same task?
For a tracked claim, exactly one writer wins. The stale claim is refused and surfaced instead of silently overwriting the current owner. That does not magically prevent overlapping files, branches, ideas, or work nobody recorded.
So it’s a thin wrapper over MCP?
MCP is the transport. The product is the write contract underneath: atomic compare-and-swap claims, leases with heartbeats so a dead agent’s work falls back into the queue, a weighted queue gated on dependency readiness, and a redaction boundary on what agents can read.
Does Sift Q listen to every conversation?
No. Ambient capture depends on a connected agent following the installed Sift Q protocol and using its task tools. The server does not independently listen to your conversations.
Does the ranking learn automatically?
No autonomous-learning claim here. Sift Q can accumulate decision evidence, while ranking changes remain deliberate, inspectable, and separately evaluated.
Who builds this?
One person, with agents — coordinated through Sift Q itself. Early access is small on purpose: design partners get a direct line, and what they hit shapes what gets built.
What is free?
Local tasks, ranking, dependencies, and lifecycle stay free. Shared coordination is the paid layer. Exact hosted pricing will be revalidated before launch rather than guessed here.