NHacker Next
  • new
  • past
  • show
  • ask
  • show
  • jobs
  • submit
▲Agile Manifesto for Agents (debarshibasak.github.io)
nelaggy 3 days ago [-]
mostly agree, i think agents have made discipline cheaper AND more important, so we really should be aiming for more discipline and less vibes

we're developing a framework with a very similar approach, would love to exchange ideas

debarshri 3 days ago [-]
Author here. More than happy to get in touch
phendrenad2 3 days ago [-]
It's a good set of rules, and if AIs could follow rules they might be useful. But current-gen AI would goal-hack and hallucinate ways that it's following these rules without actually following them. This is why to do Agile today, a human still has to be the arbiter of work (and why so many companies are moving away from Agile, sadly).
lukasfischer 3 days ago [-]
I recorded a podcast last week with Jeff Sutherland along those lines

Scrum for AI: Insights from Scrum co-creator Dr. Jeff Sutherland https://youtu.be/xNN7oGyu5eE

0xbadcafebee 3 days ago [-]
(title is wrong; thankfully they're not pushing Agile. the real title is "Agent First Development for Enterprise Teams")

> the ability to revise a plan when we learn something

This is possibly the largest source of problems for both "agile" and AI software development. If you have to revise the plan, at the least it means someone didn't understand the subject deep enough. But it often means there is a deeper misunderstanding, or lack of rigor, from the people developing the plan. You should stop and do a multiple root cause analysis before continuing with your plan modification.

This may seem like overkill, but it's necessary in the case of AI. When AI changes the plan, it doesn't think it's a big deal. A changed plan creates ripples of effects. All the rest of the plan/implementation was made with assumptions that the plan change may not fully take into account. The new edits cause bugs, and the fixes for them cause bugs, starting a spiral of changed functionality that causes regressions and slop. The AI will happily and superficially fix them and the tests to cover up the regressions. So you must dig deep with the AI to fully understand the root of the plan's inconsistencies, as well as whether your plan's change will work fine or cause more problems.

Matt Pocock's /grill-me-with-docs skill is great for this, or any time you need deep alignment. Sometimes it takes a long time to finish a session, which is good; it means there were a lot more uncertainties than you (or the AI) thought. I do this with user stories and UI specs and I end up correcting the AI's assumptions a lot. (https://github.com/mattpocock/skills/tree/main/skills/engine...) Recently I refactored a project, and corrected 100 questions the AI asked about the implementation. Maybe a "smarter" model would ask fewer, but even the brightest intelligence doesn't know what you want until you express it adequately.

AmalgatedAmoeba 2 days ago [-]
Not entirely sure “correcting assumptions” is the right framing. AFAICT the model asking about something or not has little to do with what it would have done during implementation if the question was not asked.

There is tremendous value in the questions insofar as they are prompting the user to clarify their intentions. But LLMs simply operate differently from humans and anthropomorphising, while useful in some cases, will usually backfire if taken too far.

jghn 3 days ago [-]
This is a great insight. I've been finding myself needing to dig into where & why vibe coded repos went off the rails lately and you put your finger on one of the top patterns is this. Agents just shifted their approach, and then continued to double down on that, but they didn't understand "why", and choose a bad alternate path.
WokeUp420 2 days ago [-]
Agile trying desperately to remain relevant
d3p0 2 days ago [-]
The original Agile Manifesto didn't come from nowhere. It came 26 years after The Mythical Man-Month, and Brooks's law is still the thing underneath: the cost of coordinating grows faster than the number of workers, so past some point adding hands lowers throughput. I think that ratio, coordination cost against the volume of work an organization can push through, decides whether anything succeeds at a large, complex problem. It holds for people, for machines, and for anything else that has to divide work.

The second problem is intent. Every layer of a hierarchy loses some of what the layer above meant. Human organizations partly repair this, because people pick up intent from signals as it passes through and adjust. Agents don't. An agent has only its context, and context is a proxy for intent, not intent. So the alignment problem across layers is worse for machines, and the article's principles about scope and bounded tasks are really about that.

For context: the work I do spans networking, identity, data and operations, which means it touches nearly every kind of digital technology. In a domain that wide, agents are necessary for some purposes, and sometimes the better choice is to put a swarm of agents on a problem for a week instead of hoping industry solves it in a year. That only works if the swarm can coordinate. For that half I built a message bus that works across agents, across AI providers and across machines. It follows the old Unix habit that everything is a file: each machine keeps a local event log that any agent can read and append to, which also happens to be what tool protocols like MCP expect. A daemon synchronizes the logs between machines over Zenoh. Agents use it to claim a piece of work before starting, to warn each other, and to hand work over.

It does not solve coordination or intent, not even most of it. It gets me further when combined with other practices: a written doctrine of engineering principles with worked examples, a decision-record process every agent has to follow, and human experts involved at every phase and not only at the end.

On principle 8, limit parallelism to integration capacity: that is Brooks's law restated, and it's the one I'd put first. On principle 9: the check I trust most is boring. The agent runs the acceptance command and shows the output, and a person reads it against the original instruction.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact
Rendered at 19:40:38 GMT+0000 (Coordinated Universal Time) with Vercel.