cache

A simple tokenmaxxing playbook. Briefly talk about how to increase token throughput (queues, fan-out, more context, loops) to generate high quality work (repetition, verification, evals).

Closing the software loop

• PR review agents, PR generation agent (review error logs, confirm error, reproducing error
• Issue triage agents (labeling, adding context to PRs), documentation agents (review code diffs to update knowledge), security/audit agents
• Test writing agents, refactor/migration, Q&A agents, monitoring/incident/cause grouping agents, dependency audit agents, etc.

Patterns

• Humans gate where being wrong is expensive
• Tokenmaxxing = context size x pass count x call count x frequency x io type (pdf, datasheet, diagram, text, etc.)
• Add context, world expansion, generate more things
• Multiple passes: spawn independent agent passes on the same problem, verify passes, solve same problem multiple times (fixes confidently wrong outputs, stress test pros and cons)
• Spawn many agents to create queues to drain them (multiple step research, crawlers, testing many ideas)
• Synthetic data generation: training examples, test cases, etc.
• Loops to revise (generate, critique, revise loop)
• Map-reduce over large inputs
• Throughput stoppers workarounds (e.g. rate limits, human intervention)

Tooling

• MCPs (access layer like github, CI, sentry, etc.), Claude Code, Agent SDKs
• Schedule (cron), Loop (watch)
• Skills (how to do things; pr-review, pr-generation, etc.); use evals on skills (positive and negative cases) for consistent results
• skill-creator + eval-viewer to scaffold skills and review results

Guardrails

• Eval set for each agent/skill (set expectations on correctness)
• Measure output delta with different agents and skills or without (choose best one)
• Check output / usage; check token usage and downstream metrics