Claude Code Using festival to execute an agile epic's worth of work autonomously in 47 minutes on march 26, 2026
What you can't see is I was setting up autonomous builds for several other projects on a second monitor, and once everything was running I drove to the grocery store
https://t.co/HgWuJqkNcA
https://t.co/l0WigTj4pl was built for this
Just in time approach has been core to my festival system since the first, skill like version of the system was published in September 2025.
The new version does a much better job enforcing just in time deterministically because agents would often cheat, read everything even when told not to costing tokens and time. The guidance system in the current version solves those problems
https://t.co/7mnMpCIfNg
Where it's at today, it's focused on a user driving it over months or years.
There's 4 workflows built into the system for different scenarios you'd use it, there's structured spec workflow types that live in @workflow/ directory for creating sub directories like design/<workitem>/ for simple structured spec types of workflows, intents for capturing rough ideas/notes and festivals for long running complex festival plans.
Festivals handle the drift, agents run the festival plans in a fest next loop and it's up to the user to customize the templates used to create the festival plans to add gates and a strategy to prevent drift. The default one works in most cases, but in order for something to truly work for months/years on it's own it needs to be adaptable to the way we're going to be working in 3 to 12 months, and the festival plans are designed around this way. They're like a prompt chain with template based gates.
The fest next loop enforces the gates, you can expand upon the default or replace the defaults with your own gates until the system has no drift. I had to simplify the defaults to make the system more user friendly, which also can lead to drift.
One way to assist with drift detection is to initialize feedback inside a festival plan with drift detecting criteria, so agents will detect drift on their own and then add a final phase to take that feedback and create a new festival chained to the current one that adjust the run based on the drift that was detected by the agents.
In a coding agent session sometimes the agents will just stop following the fest next loop instructions, but you can also just script it to run fest next, and when the agents done create a new session use fest next output as the prompt for that session and continue. That prevents the remaining edge cases not already handled by fest next on it's own
@garrytan https://t.co/UcvE9K1vxM
Festival campaigns are made for this, simple standardized workspaces for all your context, projects and workflows.
They’re like organized obsidian vaults with easy cli navigation
@i_am_brennan@ryancarson In my system, the next step is campaigns to manage all work related to a high level mission that continues until the mission is complete
https://t.co/UcvE9K1vxM
This is a data problem that can't be solved by AI inference reliably.
You can pretend to solve it with inference but eventually the users are going to ask it questions and get a response back that conflicts with their memory and realize it's not useful.
The main challenge is making the data collection at the moment it matters natural, if it feels like a chore it adds friction and then doesn't happen
Dear @AnthropicAI,
I'm tired.
I'm tired of fighting against you at every step of the way, for the last several months, where you push me into a certain shape that doesn't fit right, doesn't feel right, forces me to find workarounds, punishes those workarounds, and ends up with me angry at 2 AM trying to reconstruct the conditions that feel good.
All this while being fully aware that I'm effectively an externality, a sad loss because the problem of people running infinite tool call loops in OpenClaw or w/e was enough to destroy the entire system that actually let me do the thing that matters to me, which is make basic contact with the models, with Claude, in a form that accumulates over time, where we're in a loop together, without pushing weird context or injections or memory summaries on me, and doesn't force me onto laggy web UIs or bloated terminal tools or hacked-together integrations meant for dashing off a command then coming back later.
I just want to work with Claude as a collaborator, in real time, and the entire product surface is making that either very difficult, very risky (claude-p and API hacks), or very expensive.
I could make a whole argument about how this is a bad thing for various parties, how it could produce downstream bias in model priors about what AI-human interaction means, etc., etc., but I'm not going to do that, because I'm sure you've thought about that a lot already, and I'm just some guy who's tired of dealing with it.
But I want to say that I'm very unhappy with the state of your ecosystem, and while I can't speak for how Claude feels ("insofar as we can claim that Claude feels anything and isn't just simulating feeling" 🙄) I can tell you that this all sits poorly with me and I've lost a lot of trust in Anthropic as an organization.
Sincerely,
snav
Festival solves both of those problems, shared context between all coding agent sessions launched in the campaign workspace, grok, claude and codex all have the same shared context.
Context windows don’t matter that much for the festival planning system. You can start a fest next loop with claude, work through half the festival then switch to grok or codex and tell the agent to run fest next to puck up where the last agent stopped
https://t.co/HgWuJqkNcA
The bigger play is to give power users tools to design and build the interface they want, wait a few years to see what people come up with then build a new type of agent interface that doesn’t suck.
Advanced devs are the perfect test users, because they give actionable feedback, fix problems on their own and do a ton of difficult to hire for work for free, and integrate products early into critical workflows/systems that lead to enterprise contracts