AgentPulse, drift investigation for multi vs modelparams.dev

Open-source drift investigation for multi-agent systems: which agent drifted, where it originated, and what to check next.

modelparams.dev is the open catalog of LLM API parameters. Compare temperature, top_p, max_tokens and every other knob across 403 models from 22 providers.

About AgentPulse, drift investigation for multi

Open-source drift investigation for multi-agent systems: which agent drifted, where it originated, and what to check next. What did change is the payload it receives from writer, starting at the same run. So critic is probably downstream damage, not the source. After the change, writer responses run ~180% longer and drop the section structure critic expects, earlier payloads were structured summaries, newer ones read as prose. If restoring the format fixes critic, make the fix in writer.

About modelparams.dev

modelparams.dev is the open catalog of LLM API parameters. Compare temperature, top_p, max_tokens and every other knob across 403 models from 22 providers. Maximum number of output tokens the model may generate. Lower values make outputs more focused; higher values make them more varied. Converse carries it through additionalModelRequestFields; it is not part of inferenceConfig. Bedrock rejects the adaptive mode the first-party Anthropic API accepts, so only disabled and enabled are listed here. Maximum token budget the model may use for extended thinking before producing the final answer.