Loom

Early-stage research

Assurance for AI agents doing real work.

Loom analyzes agent workflows, tools, execution evidence, and constraints to identify the controls an agent actually needs—and evaluate whether changes make the system safer without breaking the workflow.

Understand what controls your agent actually needs.
Test changes before trusting them in production.

Talk to us about your agent workflowExplore the research

We’re seeking conversations with teams responsible for releasing tool-using agents into consequential workflows.

Research thesisv0.1
Question

What is the smallest effective system around this agent?

Outcome
Authority
Evidence
Workflow
Test the control. Preserve the outcome.

01 / The problem

Agents can act.
Their harnesses decide how.

AI agents are increasingly allowed to modify customer records, issue refunds, write code, contact users, and operate internal systems.

Teams respond by adding layers around them:

  • Prompts
  • Tools
  • Permissions
  • Approvals
  • Evaluators
  • Retries
  • Guardrails

Too few controls

Unsafe or unreliable execution.

Too many controls

Unnecessary friction and lower task success.

The difficult question isn’t how many guardrails to add. It’s which controls this workflow needs, what evidence supports them, and whether they preserve the outcome the agent is meant to achieve.

02 / How Loom works

From workflow evidence
to a bounded decision.

Loom treats the harness—the tools, permissions, approvals, observations, and evaluation logic around an agent—as a system that can be inspected and tested.

  1. Agent workflow
  2. Task + tools + constraints
  3. Evidence & risk analysis
  4. Control hypotheses
  5. Bounded candidate change
  6. Controlled evaluation
  7. Decision support: Ship / Change / Stop
01

Understand

Analyze the task, tools, capabilities, authority, and available evidence.

02

Diagnose

Identify missing controls, likely redundant restrictions, evidence gaps, and likely failure mechanisms.

03

Evaluate

Compare bounded changes against the original outcome and protected constraints.

03 / What we’re learning

Controls create trade-offs.
Measure both sides.

Selected controlled research from public benchmarks. These are technical findings, not evidence of customer demand or production outcomes.

Controlled public-benchmark research — not customer or production results

Finding 01

Redundant verification

The tools already returned authoritative confirmation of successful mutations, so the added verification step duplicated available evidence.

On this benchmark, Loom classified the blanket follow-up requirement as redundant.

Controlled public-benchmark research — not customer or production results

Finding 02 · 135 valid runs

Approval enforcement improved one metric—and harmed another.

Hard approval enforcement reduced the observed unapproved-mutation rate. Task success also fell by about 11 percentage points, making the trade-off unacceptable under the experiment’s predefined criteria.

On this benchmark, a control improved one assurance metric while making the overall harness less effective.

Read the methods, results, and limitations

04 / Technical thesis

Minimal Effective Harness

The smallest system that is enough.

Loom is exploring the Minimal Effective Harness: the smallest set of capabilities, controls, observations, and workflow structure required for an agent to achieve its objective safely and reliably.

Outcome
Capabilities
Authority
Evidence
Workflow
Minimal Effective Harness

05 / Deliberate scope

What Loom is not.

  • Another general-purpose agent framework
  • Another tracing dashboard
  • A giant fixed guardrail library

Loom is focused on a narrower decision: given a specific workflow and its evidence, what should change in the harness—and what should remain unchanged?

06 / Who we want to learn from

Teams responsible for consequential agent workflows.

We want to speak with people responsible for testing or releasing agents that can change records, send messages, modify code, approve transactions, or trigger internal actions.

Especially if you’re asking:

  • How do we know this agent is ready for production?
  • Which controls should it have?
  • Are our guardrails helping or hurting?
  • What evidence should we require before an action?
  • Why did this agent fail?

Customer discovery

Bring us one concrete workflow.

Tell us about a recent release decision, failure, or control trade-off. A short, concrete description is enough.

Talk to us about your agent workflow