Ship != Die

Raw cards for tools, workflows, models, infra, memory, and evals worth finding again.

22 cards query: all / type: all
agent harness

T3 Code for working across coding harnesses

A builder described using T3 Code to organize work across several coding harnesses and delegate subtasks between them. The discussion offered it as an option for people who already switch between coding apps.

T3 Codecoding agents
Group discussion
memory/state

Keep agent handoffs and verification on the issue

Teams compared ticket-based coding workflows. Recording the implementation plan, review findings, and verification results on the issue gives the next agent a place to recover the reasons behind a change.

issue trackingagent handoffsverification
Group discussion
model

Evaluate planning and implementation separately

Builders reported different preferences for writing plans and executing them. One workflow used a reasoning model to settle open questions, then handed a bounded implementation task to a faster coding model.

model selectionplanningcoding
Group discussion
review/eval

Measure what a scheduled agent actually processed

Compare incoming work with completed work. For an inbox, count messages received and handled. For a matching job, inspect a small sample of decisions. Both checks can catch failures that a successful exit code misses.

scheduled agentsquality checks
Group discussion
infra

Alert when a scheduled agent misses its deadline

A dashboard built from completed runs can look healthy after a job stops running. Record the expected schedule independently, check for a completed run at the deadline, and send one daily report of exceptions.

scheduled agentsmonitoringalerts
Group discussion
skill/workflow pack

Have a second model challenge the implementation plan

Write a ticket with testable acceptance criteria, affected files, and explicit exclusions. Ask a model from another family to look for gaps and simpler approaches before handing the revised plan to an implementation agent.

code reviewplanningacceptance criteria
Group discussion
review/eval

Frame call-reason classification as a choice

For a call-routing experiment, the group suggested giving the classifier a fixed set of possible reasons. That makes competing models easier to compare on the same calls. The discussion did not report a winning model.

classificationevaluation
Group discussion
skill/workflow pack

Separate finding subscriptions from canceling them

A discussion about assistants reviewing recurring charges raised a practical permissions question. Reading transactions, recommending a cancellation, and acting on an account each require a different level of access.

permissionsassistants
Group discussion
agent harness

Separate parallel Claude Code sessions with worktrees

The group discussed using Remote Control to start several coding sessions from Claude Desktop. Give sessions their own Git worktrees when they need to edit the same repository independently.

Claude CodeGit worktreesparallel sessions
Group discussion
skill/workflow pack

Continuous Learning Plan Generator

Directed generator pattern that builds an evolving learning plan/wiki with videos, links, and references using Tavily plus model search; expensive high-quality models are reserved for orchestration and final combining.

learning plantavilyclaude searchorchestrationcost control
curated from group research-workflow discussion
skill/workflow pack

Knowledge Base Drift PR Agents

Workflow pattern: maintain client or project knowledge bases with agents that detect drift or new insights and open PRs against markdown wikis, skills, or repo-backed context files.

knowledge basedrift detectionpull requestsskillsagents
curated from group knowledge-base maintenance discussion
memory/state

LLM OS MCP Wiki Knowledge Base

Independent wiki repo structured around Karpathy-style LLM OS, exposed through MCP and reinforced by a skill, can become an ask-all business knowledge base and context layer.

mcpwikillm osknowledge baseskills
curated from group wiki/MCP architecture discussion; personal names stripped
tool

Tavily Research Search Layer

Tavily was recommended as an initial search layer for research and learning-plan generation, with model-native search as a backup when synthesis quality matters.

tavilysearchresearchlearning planswiki
infra

LLM Serving Efficiency Patterns

Production inference checklist covering continuous batching, prefill/decode disaggregation, KV cache management, and prefix-aware routing as the real engineering layer behind model cost and speed.

llm servingcontinuous batchingkv cachegpu infraprefix routing
agent harness

Kimi Agent Swarm For Benchmarked Workflows

Agent-swarm pattern discussed for workflows that normally require human intervention: swarm the task, capture bugs and outcomes, then compare against a human benchmark to judge cost and quality.

kimiagent swarmbenchmarkingbug reportsworkflow evals
curated from group agent-swarm discussion; no external link shared
agent harness

Native Local Harness Setup Notes

Group signal that native/local harnesses are outperforming cloud-based runs for Codex in this workflow, while terminal use remains preferred for Claude. Treat harness choice as a practical efficiency lever alongside model choice.

codexlocal harnessterminal workflowclaudeefficiency
curated from group harness discussion; personal names stripped
review/eval

Claude Octopus

Claude Code plugin that runs multiple models against a coding task so blind spots show up before shipping. Most useful as review/eval support.

multi-modelclaude codereviewdebugging
model

GPT-5.5 Coding Model Note

Group signal that GPT-5.5 is being evaluated as a strong overall coding model compared with Opus 4.6 and 4.7.

gpt-5.5opuscoding modelmodel comparison
discussion note from WhatsApp group; no personal names stored