Building the new era software development lifecycle @glideapps, previously @github
Cambridge, UK
Joined December 2024
- Tweets48
- Following55
- Followers20
- Likes53
There are two things that make applied agentic workflows hard today:
1. Appropriate context for the agent
2. How to get the attention of a human
If you can solve both of these things elegantly, you can apply agents at scale effectively. Without them, you are rate-limited by humans, which a) isn't taking advantage of the technology and b) leads to questions like "is this actually serving us or making our jobs harder"
We have handed a bunch of rote engineering work over to agents. I wrote about how we do this and the numbers we are seeing in reality.
glideapps.com/blog/inside-sl…
Knowledge and context extraction from the business and humans continues to be a huge bottleneck for agent efficacy. We are working relentlessly @glideapps to solve this so we can solve it for ourselves and our customers
I haven’t written code in 18 months. I haven’t reviewed code for about 8 months.
I look back on reading/writing code with nostalgia. However, for me it was always a means to an end. It was the making of something that attracted me to software.
The future is bright for software engineers.
Unpopular opinion:
It increasingly feels like agent capability is becoming less of a bottleneck for applied AI at scale.
The harder part is everything around it: security, governance, and extracting knowledge from workers so that AI is effective
Ofc there is still a category of problems that need deep capability. But for the average everyday business task, I don't think more capability is the constraint.
One of the main challenges of building agentic workflows & loops is getting feedback from humans. LLMs are only as good as the context you give it. Without the necessary context its ability to have a shot on target is drastically diminished.
In order to recognize the value of AI at scale you have to start small (knowing how much you are leaving on the table) and the output needs to be super high signal. The loop then needs to fold the feedback back into the loop so that next time the agent has a closer shot on target.
The problem with building long-horizon agentic loops that matter is that the time to feedback is so long. It's a slog. So much so that our second generation of loops is coming with built-in evals and dry-run flow
The software engineering role as we know it is atrophying. I see two paths to replace it.
Software engineers will either:
1. Move into platform engineering to enable the rest of the organization.
2. Into a product role.
At @glideapps, bugs, security, code health, reliability, and architecture of our platform are all handled by multi-agent, long-horizon loops with minimal human in the loop. The traditional engineering toil is being abstracted away from engineers, allowing them to focus on product solutions to a stated problem. Here lies the rub. This is not an engineering role. This is what product managers have been doing for years.