Superpower Daily
NVIDIA Adds PAIR to Spread Local AI Jobs Across Nearby PCsMicrosoft Releases MAI-Transcribe-2 at 10 Cents an HourTrump Order Opens Path to Restrict Foreign Power Gear for AI Da...Google Puts WeatherNext 3’s Hourly AI Forecasts Into Search, Ma...Cohere Publishes Agent-Tool Dataset, Finds 419 Occupations Have...Uber and Wayve Put 15 Supervised Robotaxis on London’s StreetsNVIDIA Adds PAIR to Spread Local AI Jobs Across Nearby PCsMicrosoft Releases MAI-Transcribe-2 at 10 Cents an HourTrump Order Opens Path to Restrict Foreign Power Gear for AI Da...Google Puts WeatherNext 3’s Hourly AI Forecasts Into Search, Ma...Cohere Publishes Agent-Tool Dataset, Finds 419 Occupations Have...Uber and Wayve Put 15 Supervised Robotaxis on London’s Streets
Toolspublished

Oracle Publishes Agent Design That Keeps Credentials and Proof Outside the Model

Oracle’s proposed harness treats a model’s output as a request for action, leaving systems outside the model to authorize work, retain state and verify its result.

By 2 min read
Oracle Publishes Agent Design That Keeps Credentials and Proof Outside the Model
Oracle Publishes Agent Design That Keeps Credentials and Proof Outside the Model
Sign in to save and react

Listen to this story

The audio brief

About 1:38
0:001:38
Read transcript
Oracle has published a production design that keeps an AI agent’s credentials and proof of completion outside the model. The model can propose an action, but a task-specific harness decides whether that action is allowed, where it can run, and whether it actually succeeded. The case for that separation comes from a replay of 16,758 coding-agent runs. Oracle says that in two tested systems, between 60 and 69 percent of runs that ended with an incorrect answer had already found and edited the correct code. The failure was often the completion claim: the agent did the work, but did not reliably report it. Oracle’s answer is to verify completion in the system where the change occurred, rather than trusting the model’s description. Its architecture separates five layers: the model, framework, harness, runtime, and environment. The harness owns authority, containment, persistent state, and evidence. Oracle also cites results showing that this wrapper can change outcomes. On 300 SWE-bench Lite tasks, GPT-4 Turbo scored 18 percent with a codebase interface, versus 11 percent with a plain shell. Harness-Bench found a 23.8-point spread across configurable harnesses, although those comparisons do not reveal which mechanism caused the gains. Oracle packages the approach in about 300 lines of Python, using Oracle Database 26ai, SQLcl MCP Server, checkpoints, and cross-thread memory. The constraint to watch is clear: better models may need less reasoning scaffolding, but identity controls and independent verification still remain necessary for accountability.

Story brief

3 key points

Oracle’s production-agent guidance makes the harness—not the model—the accountable control plane for credentials, action permissions, state, containment, and completion evidence. The recommendation is grounded in a replay of 16,758 coding-agent runs, where 60%–69% of incorrect answers had already changed the correct code but failed to report completion reliably. Oracle also packages the approach in a roughly...

  1. 01

    Oracle separates five layers—model, framework, harness, runtime, and environment—and assigns authority and proof to the harness.

  2. 02

    SWE-agent results varied from 18.0% to 11.0% on SWE-bench Lite when GPT-4 Turbo used a codebase interface versus plain shell.

  3. 03

    Harness-Bench reported a 23.8-point spread across configurable harnesses, but the comparisons do not isolate which mechanism caused the gains.

Oracle has published guidance for production AI agents that separates reasoning from operational authority. Under its design, a model may propose an action, but the surrounding harness—not the model—holds credentials and confirms that the requested work occurred.

The distinction addresses a failure that can appear after an agent has done much of the underlying work. Oracle cites a replay of 16,758 coding-agent runs in which 60% to 69% of incorrect-answer runs in two tested systems had already found and edited the correct code. A completion claim therefore needs evidence from the system where the action took effect, rather than the model’s own account of the task.

A model proposes; the harness decides

Oracle defines an agent harness as the configured layer around a model that controls its context, permitted actions, action destinations, persistent state and proof of completion. Its five-layer architecture separates model, framework, harness, runtime and environment, assigning the harness authority, containment, continuity and proof.

That wrapper can also alter task results without changing the underlying model. Oracle cites a SWE-agent ablation across 300 SWE-bench Lite tasks where GPT-4 Turbo solved 18.0% with a purpose-built codebase interface and 11.0% with a plain shell. Harness-Bench found a 23.8-point gap between its best and worst configurable harnesses across 5,194 trajectories and 106 sandboxed tasks.

Performance gains do not identify the responsible control

The comparisons do not isolate which component produced each gain, because the tested adapters combine multiple mechanisms. Oracle also cites long-context Gemini 2.5 Pro reaching 50.8% on SWE-bench Verified with a fully observable environment and simple scaffolding. Oracle’s view is that scaffolding used to compensate for weaker reasoning may shrink as models improve, while identity and independent verification remain accountability controls.

The Oracle implementation is database-centered

Oracle presents an Enterprise Data Agent Harness workshop built around an approximately 300-line Python orchestration loop, database primitives and the SQLcl MCP Server. Its LangChain integration connects Oracle Database 26ai as a vector store, while its LangGraph integration adds OracleSaver checkpoints and OracleStore cross-thread memory. Oracle also describes a migration harness that moves a RAG corpus from MongoDB to Oracle AI Database 26ai through retrieved migration skills for sampling, schema translation, transfer, verification and reconciliation.

Sources

  1. blogs.oracle.comBuilding an agent harness that survives production

Related coverage

Continue reading