Explore Prompts

Design a cost-aware planner: assigns budgets to reasoning steps and tool calls, uses early exits, and escalates only when needed. Include a cost model and guardrails.

Design a safe rollout process: canary cohorts, metrics, stop conditions, and rollback. Include how to isolate changes (prompt vs tool vs retrieval) for attribution.

Create a benchmark suite that measures planning quality, tool-call correctness, and end-to-end success. Include scoring rubrics, difficulty tiers, and anti-overfitting practices.

Design an eval harness: deterministic replays, seeded randomness, fixed tool mocks, and artifact snapshots. Provide a folder structure and CI integration plan.

Create a safe synthetic task generation plan: avoid sensitive content, prevent leakage, and validate usefulness with human review. Include how to measure whether synthetic tasks improve real outcomes.

Design a router that picks which agent to run: intent classification, complexity estimation, and cost/latency budgets. Include fallback logic and confidence thresholds.

Define specialist agent roles with distinct prompts, tools, and constraints. Provide handoff rules and an evaluation plan to ensure specialists outperform a monolith.

Design a disagreement resolution policy: majority vote, weighted expertise, verifier override, or human arbitration. Include criteria for which method applies and how to log rationale.

Design a typed A2A message schema: intents, constraints, artifacts, and status updates. Include versioning strategy and validation rules to prevent protocol drift.

Compare multi-agent topologies (hub-and-spoke, mesh, hierarchical). Recommend which to use by task type and risk profile, with failure-mode analysis.

Create a permissions matrix: tools by scope, environment (dev/stage/prod), rate limits, and allowed parameters. Include an approval process for permission elevation.

Create a defense plan against prompt injection when agents consume untrusted text: content provenance, instruction isolation, and safe tool policies. Provide a red-team test suite.