Agentic Harness
Not an app with a chatbot. An agent harness that owns the whole stack.
Claude Code and Codex are agent harnesses for software. Barrel Brain is one for petroleum engineering. It runs its own agentic loop end to end: the model, the tools, the context, the sessions and the long-running work, all built into the app.
Open standards
AGENTS.md Agent Skills MCP
The harness, top to bottom
Every layer of the agent is built in
Orchestration
long-horizon work
Context
what the agent knows
Tools
what the agent can do
Agentic loop
think · act · observe
Tools
200+ specialised tools, out of the box
Nothing to install before the agent is useful. It can parse and edit a deck, run the simulator, read the grid, run a study, build the Excel workbook and lay out the slides: the same operations you do in the app.
Reservoir and subsurface
Decks, grids, summary vectors, studies, logs, seismic
deck_parse · grid_region_stats · summary_vector_stats · study_case_enqueue · well_las_curve_get · seismic_slab_renderRuns and processes
Launch profiles, the job queue, terminals, output
launch_add · queue_run · terminal_watch · output_pane_readCode
Scripts, queries, dashboards, the scripting API
script_code_run · query_file_run · dashboard_render · script_api_getOffice deliverables
Excel, Word, PowerPoint, PDF
excel_workbook_create · excel_chart_create · pptx_slide_add · pptx_table_add · word_content_read · pdf_readWorkspace
Files, folders, search, shell, project scaffolds
file_grep · file_update · folder_create · scaffold_create · shell_executeOrchestration
Goals, loops, todos, wake-ups, tasks
goal_set · loop_set · todo_write · wakeup_create · backlog_task_createExtend it with MCP
Connect any MCP server to add your own tools: a corporate data store, a production database, an in-house calculator. A curated registry from barrelbrain.ai makes adding one a pick from a list. It works the other way too: Barrel Brain is itself an MCP server, so Claude Code or Codex can drive it.
Context engineering
Same result. Fewer tokens. A cheaper model.
A frontier model without context gets there by trial and error, and every wrong turn costs time and tokens. When the right reference, a working example and a skill for the workflow are there to find, the agent takes the direct route, and a smaller, cheaper model can do the same job.
Without context
tryerrorsearchretryerrorrewriteretrydone gets there, slowly, on a frontier modelWith context
look upadapt exampledone faster, fewer tokens, any capable modelSkips trial and error on syntax
Reference material
An Eclipse keyword manual of more than 1,000 pages, the summary-vector catalog, and documentation for every engineering-library function and the scripting API.
Skips building from scratch
Worked examples
44 complete Eclipse decks with notes, from CO₂ and H₂ storage to thermal steam, polymer flood and multi-segment wells, plus example scripts and dashboards.
Skips rediscovering the workflow
Skills, from a skill server
A skill packages a workflow, such as reviewing a deck or setting up a study. Curated skills stay in sync with the app.
Skips asking what your team already knows
Project instructions
An AGENTS.md in the project folder tells the agent how this field and this team work: naming, units, which case is the base case.
Orchestration for long jobs
A field study isn't one prompt
It's hours of work in many steps. The orchestration layer keeps the agent working without a person present, and the session header shows the current goal, loop and watchers, so you can see what it is waiting on.
- Subagents
- Hand part of the job to a fresh agent with its own context: one per well, one per case, one for the report.
- Background jobs
- Simulator runs and long scripts run in a queue while the conversation carries on.
- Todos
- A visible plan the agent ticks off, so you can see where it is.
- Goals
- One sentence the agent works toward until it's met: “history-match every well within 10%”.
- Loops
- Adjust the case, rerun it, compare with history, for a set number of rounds. Sending a message takes the session back.
- Watchers
- The agent puts a watcher on a three-hour run and ends its turn. When the run finishes, it wakes with the result.
Every session on the record
A session records which model, tools, skills and instructions it ran with, and every file it created or edited. Weeks later you can still answer “what did the agent do, and with what?”
Fair questions
What people ask about the harness
“I have Claude Code. I can do everything in it.”
Claude Code is a harness for software. Barrel Brain is the same kind of harness, built for petroleum engineering, and Claude Code can drive it.
Out of the box Claude Code can't show you a deck, a 3D grid or a summary plot, has no tested engineering libraries, and hasn't seen the keyword manual or the example decks. Keep it, and connect it to Barrel Brain over MCP.
“Token costs will get out of hand on long jobs.”
Grounding cuts the tokens a job needs, cheaper models can do the work, and every session runs inside limits.
Most of an agent's cost is trial and error. Reference material, examples and skills remove most of it, and that same grounding lets a smaller, cheaper model do the job. Sessions run under limits on rounds per turn, turn time and loop iterations, so a job can't spin without end.
“What happens if the AI vendor changes, or we want to switch models?”
Your work is in open files, and the agent runs on any model.
Decks, scripts, queries and dashboards are plain files in your folder. Project instructions, skills and tools use open standards: AGENTS.md, Agent Skills and MCP. The harness is model-agnostic, so changing the provider changes nothing about the project.