GlassFlow builds the data infra for AI agents in production. Tares feeds agents correlated data from every system & Rius traces and debugs what prod agents do.
Hamburg: meet GlassFlow at AI.SUMMIT on Nov 24. Find us in the Startup Area at CCH.
Building with AI agents? Come tell us what you're working on.
https://t.co/soSHvsNEyY
Free online workshop, Dec 10, 17:00–19:00 CET: Building a Software Factory for Teams.
How to run a crew of AI coding agents as a team, not as a solo dev. One hour hands-on, then Q&A.
Apply (limited spots):
https://t.co/PAKfmoNPri
#SoftwareFactory#ClaudeCode
An agent's log window was cut at 1,000 lines. The line that explained the incident was 1,014.
It reasoned fine over what it saw. Most agents that fail in production have a context problem, not a reasoning problem.
https://t.co/k1EgkYENRR
#AIAgents#ContextEngineering
Wanna speak about AI software factories? I will be in Hamburg next Tue at the AI-SUMMIT 2026 organized by @hamburg_ai
Let's talk about AI software factories, their impact (we have some awesome benchmarks), and how they could transform your engineering work.
You can find our booth in the AI startup area hosted by @Oracle .
Superpowers or a software factory? Different problems.
@obra 's #Superpowers gives coding agents a disciplined workflow. A #softwarefactory adds shared decisions, long-running work and enforced permissions.
You can use both. Where does your setup break?
https://t.co/8yQMduRXVs
@zachlloydtweets #5 is where most setups stop: traces go into storage and nobody reads them.
The payoff comes when the factory reads its own traces back.
Repeated reviewer findings become rules, and loops or stalled agents get flagged while the run is still going.
@pbogosta Agreed, the fix rate is the real number. The difference is a separate reviewer that's paid to say no: it checks whether the tests would catch the bug, not just whether they pass. How it works: https://t.co/LDPmPl4mkY
1/ We ran a software factory next to plain Claude Code for a month. Same model, same $200/mo plan.
Changes shipped per day: 21.8 vs 4.3
Typing per change: 5 min vs 15
Needed a fix within 7 days: 19% vs 46%
Cost per change: ~20¢ both
🧵
@omarsar0 Agree, and it pays twice. We put a small decision model in front of an LLM step as a pre-filter: on 6 days of production logs it cut LLM calls ~80%.
The LLM only sees the cases the classifier is unsure about. Tune that threshold on your own data.
Forty identical retries isn't a bug report. It's a pattern you can catch while the agent is still running.
Same tool + same arguments above a threshold = retry loop. Alert on it live, not in the post-mortem.
How to catch silent agent failures:
https://t.co/HmCZ0C0Enl
#AIAgents
Agents in Prod is coming to London 🇬🇧
Dec 3, 18:00 at Civo Tech Junction (@CivoCloud): three technical talks on taking AI agents from demo to production, then drinks. Free, with GlassFlow and @straikerai.
Spots are limited:
https://t.co/sqkYnqhRiR
#AgentsInProd#AIAgents
@mikeydsoftware For planned work with a clear "done", yes. Same hours at the keyboard, we shipped ~3x the changes with a factory vs plain Claude Code. One-off fixes are still faster in a single session. Who checks: a reviewer agent, plus a human on releases.
Data: https://t.co/gdXw4v8b2i
@distortgeekin Agree with the core idea: stop reading the diff; check the evidence. In our software factory (https://t.co/klNrr2NpBH) the reviewer doesn't ask "do the tests pass", it asks "would these tests catch the bug". Changes needing a fix within 7 days went from 46% to 19%.
@heyprolific The other half is catching failures that return 200 OK. An agent that runs the same tool call 9 times looks green on every step. You only see it when you look at the whole trace, not single log lines.
@zachlloydtweets Agree, and "measure" is the step most setups skip. We ran a factory next to plain Claude Code for a month: 5x more changes per day, same cost, a third of the hands-on time. Same model. The difference was a shared record of decisions. How do you measure a workflow improved?
Berlin 🇩🇪, thanks for last night!
Agents in Prod covered #softwarefactory, #agentsecurity and trustworthy #AgenticAI in pharma.
Thanks to @arm1end, @ryansheldrake and Alexander Götz for the talks, adesso for hosting, and everyone who joined us.
See you at the next one!
I shared yesterday at our agents in prod meetup by @glassflowdev the impact of our software factory:
_3x more shipped changes
_34% shipped autonomously overnight, with higher-quality output (see the results in the photos).
The best part came after the talk. Sharp questions, honest debates about when a factory makes sense and when it doesn't, and a room full of people already running agents in prod.
4/ Where plain sessions still win: quick one-off fixes and fuzzy, exploratory ideas.
A factory fits when you have several projects at once and a clear definition of done.