Add two agent-era principles: autonomy follows risk, reason broadly execute narrowly - #128
Open
joshdougall wants to merge 1 commit into
Open
joshdougall wants to merge 1 commit into
joshdougall wants to merge 1 commit into
Conversation
Both are restatements, not proposals. Nothing an agent may do changes. "Autonomy follows risk" is what the ten gates already implement, and what agent-era-invariants.md summarises in one line as "agents have wide latitude inside the gates". The gates page opens with mechanics rather than rationale, so someone meeting a situation the list does not cover has a list to pattern-match against and no principle to reason from. That page admits as much: the gate list is "a floor, not a ceiling". This gives the floor a reason. It also says outright that accountability is the thing that does not scale, which the collaborator statement establishes but the principles page never repeats. "Reason broadly, execute narrowly" is the one genuinely absent idea. Nothing in operating-model/, invariants/ or workflows/ states that agents should hold wider read access than write access, though it is the assumption behind read-only review tooling and query-only data access. It is framed here as a safety mechanism rather than a concession, because narrow read access causes the failure modes the collaborator statement already names: the change that ignores an existing caching layer, or reimplements logic that exists elsewhere. Placed after "Standards lead to better code" rather than appended, so they read as part of how we work rather than a bolted-on AI section. Only "Autonomy follows risk" is added to the testable list, mapped to the gates. "Reason broadly, execute narrowly" describes how access is granted rather than something CI can check, so forcing it into either bucket would overclaim.
Deploying engineering-handbook with
|
| Latest commit: |
0c95871
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://0883514d.engineering-handbook-8f2.pages.dev |
| Branch Preview URL: | https://josh-principles-agent-autono.engineering-handbook-8f2.pages.dev |
abienkowski
approved these changes
Sep 28, 2026
kalambet
requested changes
Sep 29, 2026
|
|
||
| - **Standards lead to better code.** Use a systematic approach to ensure that code is of high quality. This involves using established practices and tools to create software that is easy to read, test, maintain, and extend. Standards should be encouraged, or enforced where necessary, through established processes such as code reviews. These standards should always be documented, and automated wherever possible. | ||
|
|
||
| - **Autonomy follows risk.** Agents have wide latitude inside the gates, and that latitude is widest where mistakes are cheap and reversible. As consequence and blast radius grow, so does the verification and approval required before an action runs. What does not scale is accountability: an engineer owns an agent-assisted change at every point on that curve. Judge an action by what it touches and what undoing it would cost, not by how difficult it was to produce. |
Member
There was a problem hiding this comment.
Super complex construct. "Latitude", "blast radius" - too many allegories.
|
|
||
| - **Autonomy follows risk.** Agents have wide latitude inside the gates, and that latitude is widest where mistakes are cheap and reversible. As consequence and blast radius grow, so does the verification and approval required before an action runs. What does not scale is accountability: an engineer owns an agent-assisted change at every point on that curve. Judge an action by what it touches and what undoing it would cost, not by how difficult it was to produce. | ||
|
|
||
| - **Reason broadly, execute narrowly.** Agents should generally have more access to observe, analyse and propose than they have to mutate. Wide read access is what prevents the expensive failure modes, the change that ignores an existing caching layer or reimplements logic that already exists three directories away, so narrowing it makes the work worse rather than safer. Write access is where the cost of being wrong actually lands, so it stays deliberately smaller than the reasoning that informs it. |
Member
There was a problem hiding this comment.
Suggested change
| - **Reason broadly, execute narrowly.** Agents should generally have more access to observe, analyse and propose than they have to mutate. Wide read access is what prevents the expensive failure modes, the change that ignores an existing caching layer or reimplements logic that already exists three directories away, so narrowing it makes the work worse rather than safer. Write access is where the cost of being wrong actually lands, so it stays deliberately smaller than the reasoning that informs it. | |
| - **Reason broadly, execute narrowly.** Agents need enough context to understand the work and avoid missing or duplicating existing solutions. Let them read relevant code and documentation, while keeping changes within the approved scope. Access to secrets and other sensitive information still requires approval. |
|
|
||
| - **Standards lead to better code.** Use a systematic approach to ensure that code is of high quality. This involves using established practices and tools to create software that is easy to read, test, maintain, and extend. Standards should be encouraged, or enforced where necessary, through established processes such as code reviews. These standards should always be documented, and automated wherever possible. | ||
|
|
||
| - **Autonomy follows risk.** Agents have wide latitude inside the gates, and that latitude is widest where mistakes are cheap and reversible. As consequence and blast radius grow, so does the verification and approval required before an action runs. What does not scale is accountability: an engineer owns an agent-assisted change at every point on that curve. Judge an action by what it touches and what undoing it would cost, not by how difficult it was to produce. |
Member
There was a problem hiding this comment.
Suggested change
| - **Autonomy follows risk.** Agents have wide latitude inside the gates, and that latitude is widest where mistakes are cheap and reversible. As consequence and blast radius grow, so does the verification and approval required before an action runs. What does not scale is accountability: an engineer owns an agent-assisted change at every point on that curve. Judge an action by what it touches and what undoing it would cost, not by how difficult it was to produce. | |
| - **Autonomy follows risk.** Agents can act independently within agreed limits. Actions that could cause more harm or are harder to undo need more checking and approval. A human remains responsible for every agent-assisted change. |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Adds two principles the handbook already relies on but never states. Both are restatements of
existing design, so nothing an agent may do changes.
Autonomy follows risk is what the ten gates already implement, and what
agent-era-invariants.mdsummarises as "agents have wide latitude inside the gates". The gates page opens with mechanics
rather than rationale, so a reader meeting a situation the list does not cover has nothing to reason
from, though that page calls itself "a floor, not a ceiling". This gives the floor a reason, and says
outright that accountability is the one thing that does not scale.
Reason broadly, execute narrowly is the genuinely absent idea. Nothing in
operating-model/,invariants/orworkflows/states that agents should hold wider read access than write access,though it is the assumption behind read-only review tooling and query-only data access.
Placed with the other how-we-work entries rather than appended as an AI section. Only the first is
added to the testable list; the second describes how access is granted rather than something CI can
check.
Issues
Closes: none.