Continuum
The line

How it works

A malformed signal enters; refined briefs leave; the queue takes them into implementation; delivery is checked against the remote. First the line, stage by stage. Then the six mechanisms that make each stage stop living in the operator.

The problem

Running more agents is the easy part.

Every one of them must be briefed, watched, transitioned and remembered, and all of that lands on the single component that does not scale: the person.

Capacity went up. The choreography went up with it, and it went up in the human. Four places it hurts:

1

The session is the wrong container

Context accumulates until the window fills, and then the understanding built inside it ends. A transcript records everything and privileges nothing: the decision weighs the same as the typos.

2

The human became the scheduler

The agent writes the code. The person opens sessions, loads context, composes prompts, watches the window, saves state, repeats. Measured: one task, eight sessions, human input almost entirely "resume" and "go ahead".

3

Declared done is not verified done

Documented scar: four tasks marked done in state files, nothing on the remote, worktrees deleted, $13.50 silently destroyed. An agent that says it finished and an agent that finished are indistinguishable without evidence.

4

The bottleneck is upstream of execution

Demand arrives malformed: a Slack thread, a feeling the release is risky, a promise made in a meeting. Turning that into a dependency-ordered set of executable tasks is hours of unassisted work, and every tool on the market starts after it.

The product

One line per signal, capture to verified delivery.

Intent Distill is the surface; the prepare governor drives it. A malformed signal enters; refined briefs leave; the queue takes them into implementation. The choreography you used to perform between sessions is the product.

Fire-and-forget when you want it: the line runs unattended end to end. Whatever deserves your answer lands in a NEEDS YOU queue and notifies you, desktop and mobile, so the line never sits idle. Answering always wins; nothing outward-facing fires without you.

CAPTURE

Everything lands in an inbox that carries no obligation: a thread, an alert, an idea on the way out the door.

REFINE

Findings surface, ambiguities are forced up and resolved by you, before execution, never mid-flight.

BRIEFS

Egested as briefs a competent stranger could execute: acceptance criteria, dependency order, workflow mode.

QUEUE

WIP-limited, dependency-ordered. How much runs at once is a number you set, not a habit you fight.

IMPLEMENT

Each task moves through gated modes; every session records its cost, turns and artifacts.

VERIFY

Delivery is checked against the remote, and what could not be read reports as a gap, not a zero.

NEEDS YOU

Only three things escalate: a product or scope decision, a fact only you know, an outward-facing effect. Everything else is arbitrated with a recorded rationale.

RECEIPTS

Every decision taken in your name is recorded: the action, the object, the criterion used, and a deep link that reopens the exact session.

THE LOOP CLOSES

A run's own findings re-enter through the same front door as any signal. What the work turned up but did not do is never lost; it queues.

Mechanisms

Six things that stop living in the operator.

None of them is judgment. Some things genuinely require it: is this the right plan, is this worth starting, is this actually done. There the work stops and waits for you. Everything else moves.

1 answers: the wrong container

The task outlives the session

The session is disposable by design; the task is not. It moves through named modes, and every transition is guarded by an artifact on disk. Progress is something a directory can prove, and the next session resumes from files, not from memory.

Task · resume-maker-html-render mode gates · artifact per transition
EXP RES PLA IMP VAL FIN
exploration.md opened RESEARCHING
research.md opened PLANNING
plan.md · todo.md opened IMPLEMENTING
validation.md will open VALIDATING
Runner · one command prompts composed · hooks deterministic
pre-hook rebase onto main
compose task state + files + standards for IMPLEMENTING
session running · turn 18 · $0.41
post-hook commit · push · task.yaml written back
gate check artifacts · open next mode
2 answers: the human scheduler

The choreography is the product

The prompt for each mode is composed, not typed: the task's state, the files that must be in front of the agent, the standards that apply. Hooks rebase before, commit and push after, write state back. Eight manual setups became one command.

3 wip is a number you set

A limit that actually holds

Work waiting to run is a queue with priorities and dependencies, resolved in topological order. How much runs at once is a number you set, not a habit you fight. Something genuinely blocked stays blocked instead of being started twice.

Queue WIP 2/3 · dependency-ordered
signals-aggregation-pass running
telegram-approval-buttons running
resume-maker-html-render queued · priority high
portfolio-site-deploy blocked · depends on ↑
gaps-study-list backlog
NEEDS YOU escalation by exception · receipts below
SCOPE DECISION · waiting on you
Split "unify auth" into two tasks (API + session store), or keep one?
arbitrated · retried flaky fetch · criterion: transient error receipt ↗
arbitrated · kept task granularity · criterion: one deliverable receipt ↗
arbitrated · reordered milestone DAG · criterion: dependency receipt ↗
4 answers: the upstream bottleneck

The refinery closes the gap upstream

Signal to findings to decisions to briefs, with adversarial lenses applied while the refinery can still fix what it wrote. Only three kinds of decision cross to you; every other call is arbitrated with a recorded rationale and a receipt you can reopen.

5 answers: unverified done

Done that proves itself

Gates are files on disk, checkable by anything that reads a directory. The delivery scoreboard reads the remote, never the local checkout, and prints the ref, the sha, and when it looked. A false "done" gets flagged, not shipped.

Delivery scoreboard origin/main · sha 91e74f7 · read 18:04Z
auth-shared-middleware delivered
auth-token-rotation delivered
auth-docs-runbook delivered-something-else
checked 8 of 9 1 unreadable: a gap, not a zero
Ledger · cost by mode per session · per task · from the file
EXPLORING $0.24
RESEARCHING $0.14
PLANNING $0.58
IMPLEMENTING $0.72
VALIDATING $0.30
task total $1.98
6 cost is a column, not a surprise

An economics ledger nobody else has

Cost and turns per session, per mode, per task. Where the time goes, which mode is expensive, whether a task was badly sized: questions with answers. A supervisor can be wrong about their own system and find out.

All product views on this page are stylized recreations of the live surfaces, with representative data.

Supervision

Supervision that never stalls

Autonomy that cannot be inspected is opacity with a better name. Every mode is named, every gate is checkable, every session can be reopened by hand, and the merge is always yours.