Skip to content
MONAD
Request access

Monad, from Engram

Superintelligence for R&D.

Not another assistant for scientists. A thinking engine for science and engineering: it plans the experiment, runs it on machines of its own, and brings back evidence you can check.

Request accessRead the thesis

The thesis

The models got good. The environment around them did not.

Research still lives in notebooks that forget, scripts with no provenance and papers nobody can keep up with. Chat returns answers. Science needs evidence.

Reasoning is becoming a commodity. Verification, provenance, memory and isolation are not, and they are what turn an answer into a result.

So we did not build another chat window. We built the environment where machine intelligence does the work of research, and shows it.

Work produces evidence. Evidence expands autonomy.

Fields

Every field has its instruments.

Monad’s agents work the way your discipline does: in its tools, against its checks, on machines of their own.

CAD & simulation

Change the geometry. Rerun the solver. Keep what holds.

Example, a session in Monad: asked to bring a flange’s peak temperature under 85 °C without adding mass, the agent reads 14 sources, sweeps fin pitch and thickness across six variants on separate machines, and reports that variant 4 reaches 83.6 °C with mass unchanged, in a change request ready for review.

  • Sweep a design parameter across variants, each on its own machine
  • Check mesh convergence before trusting a number
  • Return the variant that meets the target as a reviewable change

EDA & hardware

Every change verified in simulation before anyone reviews it.

Example, a change request in Monad: the agent adds a pipeline register on a receive path, and its summary reports worst setup slack moving from −112 ps to +38 ps, no new hold violations and 214 regression tests passing, with Dismiss and Merge changes for the reviewer.

  • Edit RTL and run the regression suite
  • Read timing and signal-integrity reports
  • Flag the change that closed the margin, or broke it

Physics & simulation

Long-running models, watched and resumed while you sleep.

Example, triggers in Monad: a nightly equilibrium scan runs 24 cases on separate machines; the agent’s session summary notes one case diverged, restarted from its checkpoint and converged, and that a change request with the scan summary is open.

  • Run field solvers and parameter scans in parallel
  • Watch long jobs and pick them up from their last saved state
  • Compare each run with the sources it set out to test

Life sciences

Literature and data, read with provenance on every source.

Example, a plan in Monad: from 22 sources the agent drafts a hypothesis, success criteria and method for a binder variant screen, then asks one question before spending compute: how many variants go to the wet lab.

  • Search papers and databases against a precise question
  • Analyse sequencing and assay data in notebooks
  • Draft the next experiment, with its controls

Software engineering

The proving ground: the tightest verification loop there is.

Example, the floor in Monad: five sessions on one service, two merged or passing, one running, one failing its tests, and one flagged because two specs disagree on a header format.

  • Many parallel sessions, one main line
  • Tests and review before anything merges
  • Abandoned attempts leave the main line untouched

Some of these suites run with partner labs today; others are being built with them. Each example is a sample project in the product; the studies and their numbers are illustrative.

The engine

One loop. Many minds.

Every field runs the same loop. Monad gives each turn of it a machine, a record and a reviewer.

  1. 01QuestionA goal in plain language. An agent reads the sources you point it at and drafts a plan: hypothesis, method, success criteria.
  2. 02ExperimentEach attempt runs on its own isolated cloud computer, in parallel, with the toolchain your field needs.
  3. 03EvidenceResults return as a change request: the diff, and the agent’s account of what it ran and found. A person reviews it.
  4. 04MemoryWhat a merged session learned stays in the project, plain and versioned, and rides into the next one.
  5. 05AutonomyAgents act inside the permissions you grant. You widen them as the record earns it.
What carries forward: a project is plain, versioned files — agents, skills, triggers and memory — that every new session starts from. Sample project.

Governance

Autonomy you can audit.

An agent is only worth trusting if you can see what it did. Every action is contained, attributed and reviewable.

Isolation
Every session runs on its own cloud computer, on its own branch. Parallel work never collides.
Review
Agent work reaches the main line through a change request. Merge is default-deny for agents; an admin can grant it.
Permissions
People and agents are both principals, with per-resource permissions and a full audit trail.
Secrets
Encrypted per project, each with an exposure you choose. Connector and model-gateway credentials are brokered server-side and never enter the sandbox.
Approval gates
Gates are off until you set them. Set them explicitly wherever you want a person in the loop.
Provenance
Every write is attributed and versioned, so the history of the research is searchable.

Questions

Before you ask.

Who is Monad for?

Research and engineering teams who want agents doing real segments of the work: simulation, analysis, hardware, software and the literature around them. Researchers get the loop in one place; R&D leaders get agent work that is reviewable and attributable.

How do I get access?

Monad is in closed beta, and access is by invitation. Request access and tell us about your work, and we will follow up.

Where does the work actually run?

Every session runs on its own isolated cloud computer, on its own branch of your project. Use the default environment or bring your own toolchain as an image. Sessions are CPU machines; heavy GPU work runs on compute you connect.

What can an agent do without asking?

What you allow. People and agents both have per-resource permissions, and merge is default-deny for agents. Approval gates are off by default: set them explicitly wherever you want a person in the loop.

How are credentials handled?

Secrets are encrypted per project, and each has an exposure. With the environment exposure, the value is available to commands in the session. Connector and model-gateway credentials are brokered server-side and never enter the sandbox.

Does it replace my CAD tools, notebooks or cluster?

No. Agents run the toolchain you package as an image, on their own machines, and reach your systems through connectors: thousands of apps through one scoped token, plus MCP, OpenAPI and raw HTTP.

Which models does it use?

Monad routes work to the model that fits the task, and you can connect your own provider keys.

How does Monad relate to Engram and Locus?

Monad is an Engram product. Engram is an applied AI research lab; Locus is its memory system. Today a project’s memory is plain, versioned files; integration with Locus is being built so that what is learned compounds further.

Engram

From an applied AI research lab.

Engram is architecting cognition: continuing Engelbart’s program of augmenting human intellect, with this decade’s tools.

Founded in 2024, with headquarters in San Francisco and R&D in Melbourne and Los Angeles.

engram.org

Work produces evidence. Evidence expands autonomy.

Bring us a hard problem.

Monad is in closed beta, and access is by invitation. Tell us about the work you would bring.

Request accessSign in