Farstax Field Notes
Farstax runs on a Farstax workspace. These notes follow the actual work of building software, operating it and testing where agents can help run a business. We record failures, decisions and unfinished experiments alongside the things that worked.
4 October 2026
Coding agents were quite capable of deciding that CI or a deployment needed time. The failure was more mundane: the turn ended, so there was nobody left to come back and check.
Read field note →5 September 2026
I kept trying to add an agent layer around coding agents that already knew how to do the work. Agent Bridge got simpler when I stopped competing with them and made the surrounding work durable instead.
Read field note →25 August 2026
Farstax's own production workspace had several generations of recent decisions compressed into one awkward migration: an old route, a proxied domain and a temporary secret bridge. Using the normal customer paths taught us more than another clean install could.
Read field note →20 August 2026
A large number looked like evidence until we asked what it counted. The lesson was not about analytics. It was about refusing to make decisions from numbers we do not understand.
Read field note →16 August 2026
We kept asking how long an agent should run without us. The better question was where one period of responsibility ends and the next has to be earned.
Read field note →13 August 2026
We want agents to stop asking humans for routine implementation choices without quietly giving them permission to create new external relationships.
Read field note →12 August 2026
Some of the clearest progress in Agent Bridge and Farstax came from removing workers, health machinery and provider paths we had previously built.
Read field note →11 August 2026
The useful part of autonomous work is continuation. The dangerous part is also continuation. Our hardest bugs are teaching us to design the stopping conditions just as carefully.
Read field note →7 August 2026
If we decide every task in advance, we can measure execution but not agency. Our current experiments start with an outcome, evidence and boundaries instead.
Read field note →5 August 2026
Farstax can see things happening in a customer workspace. We nearly let that visibility turn into control. Pulling back clarified the product.
Read field note →1 August 2026
Agent Bridge began as a small attempt to keep useful agent work moving when I was away from my desk. That changed the question I was trying to answer.
Read field note →17 October 2026
The recovery interface was showing recovery points but still leaving customers to guess what had been checked.
Read field note →17 October 2026
A fleet upgrade appeared to reach the appliance while an isolated workspace guest still had an old database.
Read field note →3 October 2026
Our weekly provider update check found newer versions of Codex, Grok, Agy and Cursor. We promoted almost none of them. “Newer exists” turned out to be only the first question.
Read field note →2 October 2026
We kept asking binary questions of systems with several independent states. A workspace could be “ready” without its runtime. A provider could fail qualification but still work. Authentication could look complete before the runtime had proved it.
Read field note →27 September 2026
We had a healthy Platform revision while lifecycle commands could still resolve stale code from /opt. It forced a clearer deployment rule: recovery has to converge on one resident owner that survives the workload it is replacing.
Read field note →23 September 2026
A deployment filled the production disk after the healthy bot had already been stopped. Then rollback failed because rollback needed disk as well. We had budgeted for the change, not for getting safely back out.
Read field note →19 September 2026
Several providers were giving us useful reasons for failures. By the time the message reached the user, those reasons had become “Internal error”. Recovery was trying to operate on information we had already thrown away.
Read field note →11 September 2026
Claude had its own permission settings and Agent Bridge had its own authority model. Both were reasonable in isolation. Together, they meant the same tool call could have two answers to the question of whether it was allowed.
Read field note →24 August 2026
The Secrets UI kept becoming awkward because we were asking it to explain a product model we had not finished. The useful fix was not another layout change. It was getting clearer about what belongs to a workspace and what belongs to an application.
Read field note →23 August 2026
Three reasonable fixes did not repair a broken repository setup because each one implemented a plausible diagnosis before we had reproduced the real failure. Agents make changing code cheap. They do not make the first explanation true.
Read field note →21 August 2026
A new autonomy path was correct in code and still took the main bot down because deployment had never learnt about its database. Finished has to include the journey into production.
Read field note →19 August 2026
Green CI felt like confidence until stale fixtures, implementation-shaped tests and a giant resource workaround showed us how easily a metric can stop measuring reality.
Read field note →9 August 2026
Models are very good at explaining why their work should have succeeded. We have become much more interested in evidence that does not depend on the model agreeing.
Read field note →17 October 2026
The useful part of extensible sensors is the observation they provide, not a new authority to change the business.
Read field note →24 September 2026
We seriously evaluated running a connector service inside every Farstax workspace. The spike worked better than expected. We still decided not to build the product around it.
Read field note →18 August 2026
We worried about runaway token use before we could even measure it properly. That was a useful reminder: controls built around guesses can become architecture before they become evidence.
Read field note →17 August 2026
The tempting answer to autonomous work was more orchestration. The useful answer was to leave judgement with the agent and make the mechanical boundaries boring.
Read field note →15 August 2026
We have working pieces, failed abstractions and better questions. That is enough to run the next experiment, not enough to declare a model for autonomous companies.
Read field note →14 August 2026
Our most useful architectural lessons have arrived as duplicate messages, broken provider output, deployment mismatches and review findings that exposed assumptions our tests had missed.
Read field note →3 August 2026
Our first instinct was to give agents company roles. The more we tested that idea, the less convinced I became that the org chart is the right starting point.
Read field note →