Asymptotelabs

Asymptote Labs (Beacon — agent-runtime observability) — Normalised enterprise agent-runtime telemetry: prompt text, assistant reasoning and responses, tool inputs/results/failures, shell commands with exit codes and durations, file operations with diff hashes and byte counts, MCP server calls, approval/policy decisions, per-user/session/model token and cost usage — captured across 14+ coding-agent harnesses (Claude Code, Claude Cowork, Codex CLI/Desktop, Gemini CLI, Cursor, Cline, OpenCode, Qwen Code, Grok Build, Devin CLI/Desktop, Factory Droid, OpenClaw Gateway) into one common event schema…

Asymptotelabs offers (Alternative, Sentiment, Reference) — Per-tenant: thousands to hundreds of thousands of events per developer-month (each agent session emits prompt, LLM-call, per-tool-call, per-file-op and command events, plus token/cost metrics and diffs); aggregate corpus across the customer base is plausibly 10^6-10^8 normalised events, with per-event size capped by configured redaction and size limits..

Agentic-LLM post-training: supervised fine-tuning on real prompt→reasoning→tool-call→result trajectories; process-reward and outcome-reward model training from approval/failure/exit-code labels; rejection-sampling and preference pairs from accepted-vs-denied tool calls; agent-eval harnesses and regression benchmarks; security/sandbox-evade taxonomy mining from denied actions and MCP call patterns; enterprise cost/latency benchmarking by model and harness; developer-productivity analytics.

The data is with 2 years of history.

Coverage spans US; Infrastructure Software; alternative, sentiment, reference; equities.