@builtbyalbert

building @SynthflowAI · running my work on AI agents and posting what actually works · berlin

 Berlin
synthflow.ai
Joined May 2021
  • Tweets3.1K
  • Following1.7K
  • Followers3.7K
442Media
i'm albert, building @SynthflowAI in berlin. most of my own work runs on AI agents now i post what actually works and what doesn't
1
7
669
everything i run on agents runs on opus at the highest effort setting. i stopped optimizing the model bill
167
best rule i've found for working with agents: reversible decisions are theirs, irreversible ones are mine
1
5
239
'ai can't do that' needs a model number.
147
by march, picking an AI model will be as exciting as picking a cloud region
134
claude goes auto mode by default next week. enterprise AI pilots still need a steering committee to change a prompt
1
2
219
my agents can build almost anything i ask for now. so the bottleneck is me: knowing what's actually needed. sky's the limit and tbh i'm struggling a bit with the ideas
1
154
an agent that asks you something it could look up itself isn't saving you time
1
150
with agents, "we'll fix it in v2" stopped meaning next year. it means after lunch
170
every company now runs a second software stack nobody planned: the internal tools its agents built last month
179
an agent that remembers yesterday beats one that's 10% smarter. persistent agents are the real launch of the month
162
best fix i found for agents that write essays: make them answer in ASD-STE100, the simplified english written for aircraft maintenance manuals
1
1
190
i barely use claude code for code anymore. it does my research, my follow-ups and the prep for every meeting
1
168
my agents finish the day's work before my first coffee. then all of it waits for me to click approve
2
252
vibe coding gets you to the demo. engineers get you to the second customer
1
3
283
the one change that made agents work for me: give them outcomes, not steps
1
177
had a bunch of codex accounts when opus 5.5 came out. tested it, cancelled all of them and moved everything to claude. quite a leap, to be honest
1
3
309
albert astabatsyan retweeted
albert astabatsyan
@builtbyalbertJun 27
The output gets better when I stop describing the task casually and start naming the work precisely. Not “help me with this.” More like: infer distill synthesize reconcile diagnose triage ground close the loop That language matters.
1
1
5
364
The output gets better when I stop describing the task casually and start naming the work precisely. Not “help me with this.” More like: infer distill synthesize reconcile diagnose triage ground close the loop That language matters.
1
1
5
364
come on ....
Sakana AI
@SakanaAILabsJun 22
Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API. Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls. Try it: sakana.ai/fugu 🐡
537
Load more