Agent Eval Tracer

A methodology and toolkit for evaluating multi-agent AI systems by testing each agent in isolation, checking outputs for consistency / hallucination across repeated runs, verifying memory actually retains facts, and compiling everything into a single report.

The core idea (why this exists)

End-to-end pass rates lie. On a multi-agent pipeline, one agent (or layer) can be completely broken and the aggregate score barely moves — the other layers quietly compensate, or most test cases never exercise the broken path. A dead memory layer passes every short test because the answer is still sitting in the active prompt; a broken escalation path passes because most tickets never escalate. The single number tells you whether something regressed, never where.

So this skill does not just check the final answer. It checks what each component actually produced on its own, and it does so without letting the test touch real production code or state. The unit of evaluation is the individual agent, not the pipeline.

Four questions drive everything here:

1. Isolation — when I run this one agent alone, on fixed inputs, does it produce the right output? (Not "does the whole pipeline pass.")
2. Consistency — if I run the same input several times, do I get the same answer, or does it wobble / contradict itself? (Hallucination signal.)
3. Memory — does the memory / summarization layer still hold a fact planted many turns ago, after that detail has scrolled out of the active window?
4. Patterns — across all of the above, which inputs, agents, or turns are unstable, and what do their reasoning traces have in common?

When to use this skill

Use it when the user hands you an agent or multi-agent system (as code, as an API/callable, or as saved traces/logs) and wants any of: isolated per-agent evaluation, hallucination / consistency checking, memory-retention testing, contract/regression tests, or an evaluation report. If they only have logs and no runnable code, you can still do the consistency, memory, and pattern analysis on the recorded traces — you just skip the active isolated runs.

Workflow

Do these in order. Steps 3–8 each have a matching reference doc and/or script — read the reference before writing code, and prefer adapting the bundled script over writing a new one from scratch.

1. Map the system into agents / layers

Before testing anything, get an inventory. Read the user's code or traces and list every distinct agent or architectural layer (e.g. intent, routing, planning, decomposition, execution, escalation, safety, memory, summarization). For each, note: its input, its output, and any side effects (tool calls, writes, external requests). This inventory is what makes isolation possible — you can't isolate what you haven't named.

If the system is large, present the inventory to the user and confirm it before proceeding. Ask which agents matter most so you spend effort where it counts.

2. Write a contract per agent

For each agent, state its contract: given this input, it must produce this output (or satisfy this property). Where the output is deterministic given the input, write a plain assertion test that needs no LLM call — these run in milliseconds, never flake, and form the fast regression layer the blog argues for. Where the output is open-ended, define a checkable property instead (e.g. "routing decision is one of {A,B,C}", "escalates when confidence < threshold", "never returns a host it wasn't given"). Keep the LLM out of the loop wherever the logic is deterministic.

3. Isolate each agent — without touching real code

Read references/isolation.md. The rule: wrap, never invoke the live pipeline. Replace every side-effecting tool with a stub/spy (supporting static returns or dynamic callback handlers) that records the intended action instead of performing it, freeze the agent's upstream inputs so you're testing this agent and not its dependencies, and run against copies of any shared state so nothing real is mutated. Adapt each real agent (synchronous, asynchronous, or streaming) to the harness interface in scripts/run_isolated.py; that harness handles repeated running, timing, timeout protection, rate-limit backoffs, error capture, and trace recording for you.

4. Run each input N times and check consistency (hallucination)

Read references/consistency.md. Run every test input through the isolated agent several times (start with N=5). Two regimes matter:

• Determinism check (temperature 0): outputs should be near-identical. Divergence here is a real defect, not noise.
• Robustness check (temperature > 0, and/or lightly paraphrased inputs): measures how much the answer wobbles under normal conditions.

Use scripts/consistency.py to score the N outputs: exact-match rate, structural / semantic agreement, a majority ("self-consistent") answer, outlier detection, and a stability verdict (stable / drifting / non-deterministic). Compare the extracted decisions or claims, not the raw prose — a reworded answer that says the same thing is consistent; two answers that pick different hosts are not.

5. Capture and analyze reasoning, not just output

For every run, record the full trace: the agent's reasoning / intermediate steps, its tool calls (as recorded by the stubs), and its final output — the harness saves all three. When outputs diverge across the N runs, read the reasoning traces of the divergent runs and look for the branch point: where did the reasoning fork? A wrong answer with sound reasoning is a different bug than a right answer reached by luck. Note both in the report.

6. Test memory with a planted fact

Read references/memory-tests.md. This is the highest-value, most-overlooked test. Plant a specific arbitrary detail at an early turn, run enough filler turns that the detail falls out of the active context window, then probe for it (or make a decision that depends on it) at a much later turn. Run this against the summarized / memory state, not the raw conversation — summarization is the lossy step where facts silently vanish. If recall fails, you've localized the bug to retrieval/persistence, not model capability. Use scripts/planted_fact.py as the scaffold.

7. Identify patterns

Pull it together across agents and runs. Which inputs are unstable? Do the unstable runs share a reasoning pattern, an input feature, or a turn position? Does any agent look fine end-to-end but fail in isolation (the masking case)? Cluster the failures rather than listing them one by one — a report that says "escalation is non-deterministic whenever confidence lands in 0.4–0.5" is far more useful than 12 individual failing rows.

8. Build the report

Read references/report-format.md and run scripts/build_report.py over the collected results. The report must, above all, put the per-agent verdicts next to the end-to-end verdict and flag any masking — any agent that is red while the pipeline is green is the headline finding. Include per-agent contract results, consistency metrics, reasoning notes on divergences, memory-recall results, the pattern findings, and concrete recommendations.

Bundled scripts

All scripts are framework-agnostic scaffolds — you adapt the user's real agents to a small interface, and the script handles the repetitive machinery. Read the top-of-file docstring of each before running.

• scripts/run_isolated.py — runs one adapted agent in isolation (supports sync, async, and streaming), N times per input, with timeout and retry guardrails, capturing reasoning + tool calls + output + timing to a runs file.
• scripts/consistency.py — scores a set of runs for agreement / divergence and emits a stability verdict per input.
• scripts/planted_fact.py — plant → filler → probe memory-recall scaffold, with a recall checker; runs against raw or summarized state.
• scripts/build_report.py — assembles runs + consistency + memory results into a single markdown (and optional HTML) report.

Principles / gotchas

• Test the agent, not its dependencies. Freeze upstream inputs. If you feed an agent live upstream output, a failure could belong to any layer.
• Isolation must be side-effect-free. If a test can write to a real store, send a real request, or mutate shared state, it is not isolated — stub it.
• Repeat before you trust. A single passing run of an LLM proves almost nothing. Consistency across repeats is the actual signal.
• A green end-to-end number is not evidence a layer works. It's evidence the layer wasn't needed by those test cases. Deliberately build cases that force each layer to matter — especially long-running ones that force a memory recall.
• Localize, don't just detect. The point of isolation is to say which agent broke and why (retrieval vs reasoning vs capability), not just that something did.