You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
docs: define lights-sparse and graduated in VISION
Add self-contained terminology sections so readers need not infer
meaning from the dark-factory analogy or external search.
Co-authored-by: Cursor <cursoragent@cursor.com>
Copy file name to clipboardExpand all lines: docs/design/VISION.md
+14-1Lines changed: 14 additions & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -9,7 +9,20 @@ This document states the long-term direction of **ABCA (Autonomous Background Co
9
9
10
10
ABCA is a **reference sample** of a self-hosted **software dark factory** on AWS: a platform that accepts high-level engineering intent, runs autonomous coding work in isolated cloud environments, and returns **reviewable outcomes** (pull requests, review comments, validation evidence) under human governance — not a chat session that happens to edit files.
11
11
12
-
We are building toward **lights-sparse**, **graduated** autonomy — the implementation loop is increasingly unattended while governance and release stay supervised (see the [dark factory introduction](../src/content/docs/index.md)): the default experience is **fire-and-forget** (submit, walk away, review the outcome), while the platform can **escalate to humans** when policy or risk requires it — without turning every task into a live pair-programming session. Customers should be able to **start conservative and loosen over time** (per repo, per policy, per task) as trust in outcomes grows. The platform’s job is to surround the agent with a **durable control plane** — admission, orchestration, memory, policy, HITL gates, cost limits, and observability — so that background agents are **reliable, bounded, and improvable** at organizational scale.
12
+
We are building toward **lights-sparse**, **graduated** autonomy (defined below): the default experience is **fire-and-forget** (submit, walk away, review the outcome), while the platform can **escalate to humans** when policy or risk requires it — without turning every task into a live pair-programming session. Customers should be able to **start conservative and loosen over time** (per repo, per policy, per task) as trust in outcomes grows. The platform’s job is to surround the agent with a **durable control plane** — admission, orchestration, memory, policy, HITL gates, cost limits, and observability — so that background agents are **reliable, bounded, and improvable** at organizational scale.
13
+
14
+
### What "lights-sparse" means
15
+
16
+
**Lights-sparse** is project vocabulary (not general industry jargon): it names the autonomy posture ABCA targets today, drawn from the **software dark factory** analogy in the [introduction](../src/content/docs/index.md).
17
+
18
+
-**Lights-out** (the analogy’s end state): humans set goals, policy, and constraints; production runs without people on the floor.
19
+
-**Lights-sparse** (where teams are now): the **implementation loop** — edit code, run tests, open pull requests — is increasingly **unattended**, while **governance, merge authority, and production release** stay **supervised**. Humans are not at the keyboard for every step; they are still accountable for what ships.
20
+
21
+
ABCA is built for that posture: asynchronous tasks, policy-gated escalation when risk requires a person, and a control plane that bounds blast radius and cost. The longer-term direction is **self-evolving** delivery (tenet 9): the platform learns from outcomes—memory, review feedback, evaluation—so repeat work gets cheaper and more reliable, not just one-off autonomous runs.
22
+
23
+
### What "graduated" means here
24
+
25
+
**Graduated** autonomy is not a single on/off switch: operators tighten or loosen gates over time (Cedar policies, `--pre-approve`, per-repo posture) without forking the deployment. See tenet 2 and [CEDAR_HITL_GATES.md](./CEDAR_HITL_GATES.md).
13
26
14
27
Success looks like teams that can **submit work and walk away**, trust that doomed work fails fast and cheaply, inspect every important decision in an audit trail, and see **measurable improvement** over time (fewer revision cycles, higher first-review merge rates, predictable cost).
Copy file name to clipboardExpand all lines: docs/src/content/docs/architecture/Vision.md
+14-1Lines changed: 14 additions & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -13,7 +13,20 @@ This document states the long-term direction of **ABCA (Autonomous Background Co
13
13
14
14
ABCA is a **reference sample** of a self-hosted **software dark factory** on AWS: a platform that accepts high-level engineering intent, runs autonomous coding work in isolated cloud environments, and returns **reviewable outcomes** (pull requests, review comments, validation evidence) under human governance — not a chat session that happens to edit files.
15
15
16
-
We are building toward **lights-sparse**, **graduated** autonomy — the implementation loop is increasingly unattended while governance and release stay supervised (see the [dark factory introduction](/architecture/index)): the default experience is **fire-and-forget** (submit, walk away, review the outcome), while the platform can **escalate to humans** when policy or risk requires it — without turning every task into a live pair-programming session. Customers should be able to **start conservative and loosen over time** (per repo, per policy, per task) as trust in outcomes grows. The platform’s job is to surround the agent with a **durable control plane** — admission, orchestration, memory, policy, HITL gates, cost limits, and observability — so that background agents are **reliable, bounded, and improvable** at organizational scale.
16
+
We are building toward **lights-sparse**, **graduated** autonomy (defined below): the default experience is **fire-and-forget** (submit, walk away, review the outcome), while the platform can **escalate to humans** when policy or risk requires it — without turning every task into a live pair-programming session. Customers should be able to **start conservative and loosen over time** (per repo, per policy, per task) as trust in outcomes grows. The platform’s job is to surround the agent with a **durable control plane** — admission, orchestration, memory, policy, HITL gates, cost limits, and observability — so that background agents are **reliable, bounded, and improvable** at organizational scale.
17
+
18
+
### What "lights-sparse" means
19
+
20
+
**Lights-sparse** is project vocabulary (not general industry jargon): it names the autonomy posture ABCA targets today, drawn from the **software dark factory** analogy in the [introduction](/architecture/index).
21
+
22
+
-**Lights-out** (the analogy’s end state): humans set goals, policy, and constraints; production runs without people on the floor.
23
+
-**Lights-sparse** (where teams are now): the **implementation loop** — edit code, run tests, open pull requests — is increasingly **unattended**, while **governance, merge authority, and production release** stay **supervised**. Humans are not at the keyboard for every step; they are still accountable for what ships.
24
+
25
+
ABCA is built for that posture: asynchronous tasks, policy-gated escalation when risk requires a person, and a control plane that bounds blast radius and cost. The longer-term direction is **self-evolving** delivery (tenet 9): the platform learns from outcomes—memory, review feedback, evaluation—so repeat work gets cheaper and more reliable, not just one-off autonomous runs.
26
+
27
+
### What "graduated" means here
28
+
29
+
**Graduated** autonomy is not a single on/off switch: operators tighten or loosen gates over time (Cedar policies, `--pre-approve`, per-repo posture) without forking the deployment. See tenet 2 and [CEDAR_HITL_GATES.md](/architecture/cedar-hitl-gates).
17
30
18
31
Success looks like teams that can **submit work and walk away**, trust that doomed work fails fast and cheaply, inspect every important decision in an audit trail, and see **measurable improvement** over time (fewer revision cycles, higher first-review merge rates, predictable cost).
0 commit comments