Skip to content
GEODE

Menu

Autonomous Agent Harness

Execute the task. Explore a better way to run it.

GEODE chooses its next action from tool observations. An outer Experimental Loop searches scaffolds and evaluation scenarios, then manages candidate acceptance and reversion.

Install GEODEExplore the research

Autonomous Agent Harness

Run the task at hand

  1. Decide
  2. Act
  3. Observe

↶ Use observations to choose the next action

Execution records inform evaluation and search

Experimental Loop

Test candidates for future runs

  1. Propose
  2. Evaluate
  3. Keep / revert

Separate verdicts for search acceptance, validation, and release

An implemented path from runtime observations to scaffold search. Research and evidence

Autonomous Agent Harness

A recorded TLS certificate task.

Follow the TLS certificate task from the request through tool calls to its verifier result.

How autonomous execution works
Recorded public run

Create a TLS certificate

openssl-selfsigned-cert

GPT-6 Astra / highOne task, one trial

  1. terminal_exec20:29:32 UTC
    ok
  2. terminal_exec20:30:27 UTC
    ok

2 calls matched to 2 results. Command and result bodies are withheld.

3 rounds12 public eventsRaw trajectory

Public event metadata. 9 payload bodies withheld; not a full replay.

The result becomes the next step.

The loop runs a tool, reads its result, then chooses another action or a final response.

  1. 1Your taskGoal and environment
  2. 2Choose an actionModel decision
  3. 3Run a toolPermitted action
  4. 4Read the resultContext for the next decision

Use each observation to continue the tool-and-decision loop.

Approval policies, execution budgets, and cancellation govern the run.

Explore the runtime

Harbor × Terminal-Bench 2.1

Same model. Different harness.

GEODE and native Codex, measured on the same tasks and verifiers.

Both runtimes used gpt-5.6-sol, max reasoning, and the OpenAI subscription route on Terminal-Bench 2.1. Harbor 0.22.0 supplied the same tasks and canonical verifiers. The comparator was native Codex CLI 0.145.0.

445 planned − 10 environment-unavailable − 6 unmatched = 429 common valid trialsThe 6 infrastructure-invalid native Codex trials and their GEODE counterparts are excluded. Valid task failures remain in the denominator. Bar scale: 0-100%.
RuntimePassed / validPass rate
GEODE339/42979.02%
native Codex331/42977.16%

The task-balanced difference was +1.26 percentage points; the 95% task-bootstrap interval was -5.40 to +8.05 percentage points.

A local paired-runtime comparison over the common valid trials. The 95% interval includes zero; the full-suite decision is inconclusive.

Measurement setup and accounting

Measured configuration: b549f3e thin AgenticLoop adapter, custom system_prompt_override, and terminal_exec-only tools. The preregistered full-suite primary is not-measurable, the decision is inconclusive, and official leaderboard submission eligibility is false.

The protocol registered 89 tasks with 5 repetitions, or 445 trials per runtime. bn-fit-modify and tune-mjcf were excluded symmetrically because their amd64 verifiers could not run on the arm64 host. 6 native Codex trials remained infrastructure-invalid after their supplement caps.

Infrastructure-invalid trials are excluded; valid task failures remain in the denominator. Runs spanned several days without control of shared seeds, timing, provider capacity, or credential-principal load.

Run it. Measure it. Improve it.

core runs tasks, evals measures their results, and evolve searches scaffold candidates.

core

Where the work happens

Terminal, channel, and MCP entry points use the same AgenticLoop. The runtime owns tools, sessions, and execution controls.

ArchitectureContextSkills
evals

Evidence for a judgement

Task results, trajectories, and verifier receipts support comparison and audit.

Read the evaluation method
evolve

Scaffold candidates, selected through experiments

Mutate and evaluate scaffold candidates, then keep or revert them.

Explore the outer loop

Research direction · RSI

Test what the next run inherits.

Improvement-execution controls (L1) and bounded candidate search (L2) are implemented. Next: establish whether runs inheriting accepted changes outperform a fixed-strategy control.

Further research covers experience acquisition (L3), deployment adaptation (L4), and recursive inheritance (L5).

Current scope, terms, and next testsDesign guide: RSI whitepaper v2

Start in your environment.

Install the stable release, connect a model, and give GEODE its first task.

Recommended. Install the latest stable release in an isolated tool environment.

uv tool install geode-agent

After installation, run geode. Inside the session, use /login to review credentials. Choose an authentication path

Model connectionsAnthropicOpenAI / CodexOpenRouterZhipuAI GLMChoose an auth path

Subscription authentication and metered APIs are separate routes. Model access and billing depend on your account and provider.