Tool-Using Agent Reliability Chain

Agent Plan

Chain length 4 tools

Per-tool success rate 85%

Retry budget per step 1

Reliability layer

Each observation is checked against the tool's schema before the plan advances — ungrounded failures caught early instead of compounding.

Live run

Predicted success

–

Observed success

–

Avg tool calls / task

–

Tasks run

–

How it works

An agent's plan is a chain of tool calls: planner → tool 1 → observe → tool 2 → observe → … → done. Each call succeeds independently with probability p. If a step fails it may retry, up to the retry budget r; the step only kills the task once every retry is exhausted. A grounded validator adds a schema/precondition check after each observation, catching malformed results before they poison the next step — modelled here as a flat reliability boost on top of the raw tool success rate.

P(step succeeds) = 1 − (1 − p)^(r+1)
P(task succeeds) = P(step succeeds)^chainLength

Chain length — how many tools the plan strings together end to end. Longer plans do more, but every extra link multiplies in another chance to fail.
Per-tool success rate — reliability of a single call to a single tool (a search API, a code executor, a database lookup…).
Retry budget — how many times the executor re-attempts a failed step before giving up and forcing a replan. Retries buy reliability at the cost of latency and API spend.
Grounded validator — typed I/O checks and preconditions on each tool's schema; toggle it off to see how quickly un-grounded chains degrade.

Watch the pipeline: agent tokens flow left to right through each tool node, flash red and loop back on a failed attempt (retrying in place), and either reach the goal node (green) or fall off the chain (dark red) once retries run out. The stats panel compares the closed-form prediction above against the empirical success rate measured live.

Agent token Tool call succeeded Retrying Task failed

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.