<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rajan-2026-traceforge">
    <titleInfo>
        <title>TraceForge: A Human-Gated System for Auto-Hardening Tool-Using LLM Agents into Immutable Workflow DAGs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Santhiya</namePart>
        <namePart type="family">Rajan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Sweta</namePart>
            <namePart type="family">Agrawal</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Alberto</namePart>
            <namePart type="family">Abad</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-461-3</identifier>
    </relatedItem>
    <abstract>Tool-using language agents are flexible but non-deterministic: an orchestrator LLM re-plans on every request, adding latency, cost, and paths that are hard to audit. We demonstrate TraceForge, an on-premise system that observes how an agent actually solves a recurring task and, on operator request, hardens it into a deterministic, immutable workflow DAG. TraceForge attaches to the OpenTelemetry traces an agent already emits (zero code changes), reconstructs each run’s execution DAG, and runs a human-gated loop: a statistical aggregation gate (proposes a DAG only once enough runs agree), a Judge-LLM reflection step, and a human freeze gate that writes a versioned, immutable specification with lineage and rollback. The pipeline is topology-agnostic—applied to single-agent tool-callers and multi-agent orchestrators alike—and all components run on-premise with no external dependencies. On held-out requests, frozen workflows reproduce the dynamic agent’s outputs exactly while removing every planner-LLM call and cutting latency 1.39x. The entire loop is driven from a browser console with no coding, and is shown end-to-end in the accompanying screencast.</abstract>
    <identifier type="citekey">rajan-2026-traceforge</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-demos.37/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>369</start>
            <end>379</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T TraceForge: A Human-Gated System for Auto-Hardening Tool-Using LLM Agents into Immutable Workflow DAGs
%A Rajan, Santhiya
%Y Agrawal, Sweta
%Y Abad, Alberto
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-461-3
%F rajan-2026-traceforge
%X Tool-using language agents are flexible but non-deterministic: an orchestrator LLM re-plans on every request, adding latency, cost, and paths that are hard to audit. We demonstrate TraceForge, an on-premise system that observes how an agent actually solves a recurring task and, on operator request, hardens it into a deterministic, immutable workflow DAG. TraceForge attaches to the OpenTelemetry traces an agent already emits (zero code changes), reconstructs each run’s execution DAG, and runs a human-gated loop: a statistical aggregation gate (proposes a DAG only once enough runs agree), a Judge-LLM reflection step, and a human freeze gate that writes a versioned, immutable specification with lineage and rollback. The pipeline is topology-agnostic—applied to single-agent tool-callers and multi-agent orchestrators alike—and all components run on-premise with no external dependencies. On held-out requests, frozen workflows reproduce the dynamic agent’s outputs exactly while removing every planner-LLM call and cutting latency 1.39x. The entire loop is driven from a browser console with no coding, and is shown end-to-end in the accompanying screencast.
%U https://aclanthology.org/2026.emnlp-demos.37/
%P 369-379

Markdown (Informal)

[TraceForge: A Human-Gated System for Auto-Hardening Tool-Using LLM Agents into Immutable Workflow DAGs](https://aclanthology.org/2026.emnlp-demos.37/) (Rajan, EMNLP 2026)

• TraceForge: A Human-Gated System for Auto-Hardening Tool-Using LLM Agents into Immutable Workflow DAGs (Rajan, EMNLP 2026)

ACL

• Santhiya Rajan. 2026. TraceForge: A Human-Gated System for Auto-Hardening Tool-Using LLM Agents into Immutable Workflow DAGs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 369–379, Budapest, Hungary. Association for Computational Linguistics.