Just as true in the enterprise. We’ve created so many problems (auth, security, side effects) for ourselves by trying to give everything to super agents
Stop waiting for one super agent. Every app on your phone is becoming an agent. The future is polyagentmorous. There will not be one agent to rule them all.
163
Dan Adler retweeted
it looks like cursor has stopped using vector-based code retrieval and moved away from turbopuffer
22
17
465
299,613
Under-discussed aspect of yesterday's launches (Opus 5.5 and GPT-6) is how any agent startups that price on consumption + markup will see huge chunks of revenue disappear when they bump to these models.
For now, it's okay, because demand growth is so high. but at some point, how do you deal with the market expecting a 30–50% price reduction every 3 months?
Another example of how impressive this technology really is: The cost of achieving a given level of AI performance is falling roughly 13-fold per year, according to Epoch AI.
Its analysis estimates a 47% quarterly decline since 2023, faster than any other transformative technology in history.
One example: OpenAI’s o3 exceeded 25% on FrontierMath Tiers 1–3 at a reported cost of about $0.55. Around 18 months later, GPT-5.6 Luna reached the same threshold for roughly $0.0015.
The fastest price declines tend to come immediately after a new capability appears. Across five benchmarks, Epoch estimates a 66% quarterly decline for newly achieved frontier performance, slowing to 32% two years later.
The consequence? In just a few more months, every person on this planet will have access to Fable 5-level entertainment at a fraction of today's cost. And this trend will probably continue indefinitely. Super exciting!
221
When you get high enough up in your career, whatever it is, your job will become sales
We believe there has never been a better time to get into sales. AI has made the role even more valued and in demand. We're building a new program to bring the next generation of ambitious talent into sales, but today we're also sharing an essay that outlines sales fundamentals.
The Craft of Sales: firstround.com/sales-craft
11
528
Dan Adler retweeted
Outcome-based pricing for code
We launched Agentic Batch Changes this week: the very first coding agent with outcome-based pricing. You only pay when you merge a PR, not per token.
Despite the innovation in other verticals—Sierra, Fin, Decagon, and more all charge per resolution—dev tools have remained firmly tied to seat- and usage-based pricing.
So why are we doing this? We’re confident in the product, and pricing on outcomes is the clearest, most tangible way to prove it. This pricing puts our incentives in lockstep with yours.
sourcegraph.com/blog/agentic…
1
4
30
4,147
We launched Agentic Batch Changes this week: the very first coding agent with outcome-based pricing. You only pay when you merge a PR, not per token.
Despite the innovation in other verticals—Sierra, Fin, Decagon, and more all charge per resolution—dev tools have remained firmly tied to seat- and usage-based pricing.
So why are we doing this? We’re confident in the product, and pricing on outcomes is the clearest, most tangible way to prove it. This pricing puts our incentives in lockstep with yours.
sourcegraph.com/blog/agentic…
9
6
31
24,191
And read more about the Agentic Batch Changes product here sourcegraph.com/blog/introdu…
1
3
290
As much as I like to believe that this is the future of the industry, there is a reason we can do this while others cannot. Most coding agents were designed to be general-purpose problem solvers, just as likely to be used for vacation planning or open-ended exploration as executing a code change. That makes ROI an impossible question.
Agentic Batch Changes is a special-purpose agent, scoped deliberately in its harness, its system prompt, its permissions, and its tools, to solve one concrete problem. We built it to minimize cost at scale, rather than paying for an agent to do the same job a hundred or a thousand times over.
1
199
Today @sourcegraph’s Agentic Batch Changes, the frontier agent for large-scale code changes and migrations, goes live. The Batch Changes engine already helps our customers merge half a million PRs per year for security and remediation, dependency upgrades, and ops migrations. With this release, we are bringing a frontier agent to oversee changes and adapt to the unique requirements of thousands of repos at once.
sourcegraph.com/blog/introdu…
1
6
12
633
Even after all these years... Even though I run an AI company! I still think that this is the way that AI has most transformed my day to day life.
I think one of the advantages AI has for helping you with a problem over a human is there is no shame in saying "I don't get it" over and over and over until it explains in a way that you do
1
1
160
1/ I wrote about the future of owning a codebase!
Today, the vast majority of most devs' experiences with agents start with a prompt. But since little of the work to be done in a mature codebase is greenfield features, that's turned us into meat proxies.
“Investigate this alert I got”
“check if our codebase is at risk from this supply chain attack”
“bump spring boot to 4.1.1”
The autonomous codebase
Prompt to PR might be solved, but codebase maintenance is not!
leaddev.com/software-quality…
1
1
12
1,306
4/ But I want to emphasize my “narrowly-scoped” comment. This is an unpopular take, but the challenges we face with agents in the enterprise (auth, permissions, and more) are largely our own creation. If your goal is to build and sell a single agent that can handle any prompt, you’re going to design a machine capable of anything: unlimited access, unlimited risk.
The future looks a lot more like narrowly-scoped agents that have the ability to use judgment in their domain.
1
49
5/ I’m excited to talk about our vision of the autonomous codebase on stage at @TheLeadDev NYC in 2 weeks! See you there!
45