An agent harness is the software and working rules around an AI model: how it receives information, uses tools and carries work forward. Consider a fictional supplier amendment proposing higher fees and different renewal terms. The example below sets out proposed design requirements for assessing the change. The linked research supplies engineering ideas whose value in legal work needs its own testing. Sources reviewed 14 September 2026.
Assemble the relevant context.
ReAct explores agents that alternate reasoning and actions, obtaining further information through tools. ReAct: paper
The amendment needs to be read alongside the agreement, earlier amendments, reviewed advice and current business requirements. Versions and document status matter. A missing schedule or unresolved conflict should remain visible.
An agent could request further evidence within its permitted scope. The assessment should identify what was considered, what remains unavailable and which conclusions depend on resolving the gap. Finding a document also leaves the question of whether its contents support the proposed interpretation.
Put exact checks in tools.
A defined calculation tool can work out revised annual charges using verified inputs. Software can check that required fields are present and cited passages occur in the identified document version.
Those checks answer specific questions. Correct arithmetic still depends on the right inputs and applicable rules. A matching quotation still requires an assessment of its meaning. In the supplier example, the calculation and its assumptions should accompany the recommendation so the reviewer can examine both.
Connect authority to the proposed action.
Preparing an assessment, approving a recommendation and accepting the supplier’s terms carry different responsibilities. The workflow should identify who can perform each step and what evidence they need.
Approval should attach to the actual content and action presented. A material change requires renewed review. Record the source versions, tool outcomes, checks, decisions and confirmed actions, with appropriate access and retention. These records give responsible people a basis for investigating failures. Confirm an action through its actual result before treating it as completed.
Retain knowledge people can use.
MemGPT studies movement between memory tiers to support document analysis and conversations beyond a model’s immediate context. MemGPT: paper
After review, preserve the accepted position, its source basis, assumptions, unresolved questions and responsible person. If another amendment arrives, the organisation should be able to identify which earlier conclusions need reconsideration.
For legal work, temporary working context and maintained records need distinct roles. Give each retained assessment an identified status and owner. An agent’s observation remains a proposal until it has received the review needed for its intended use.
Measure improvement against the work.
Meta-Harness reports automated improvements to harness code in classification, mathematics and coding. These results support experimentation; legal workflows require their own evaluation. Meta-Harness: paper
Compare the proposed workflow with the current process using the same representative material. Assess material omissions, unsupported conclusions, source coverage and the effort needed to review the result. Include incomplete documents and conflicting amendments.
When a failure matters, retain an appropriately permitted example and test proposed corrections against it and other cases. Keep some evaluation examples separate from those used to develop the change. Agents can propose improvements; adoption requires relevant evidence and responsible review. Preserve the earlier method so it remains available if the change disappoints.
Preserve continuity as providers change.
Anthropic describes a context-management workaround that became unnecessary with a newer model, and an architecture separating durable records, the harness and execution tools. Anthropic: Scaling Managed Agents
A change of model should preserve usable access to the organisation’s sources, reviewed decisions, working instructions and evaluation examples. Provider terms, tool behaviour and data handling also need assessment. Test the replacement against the work before relying on it.
The harness itself should remain open to change. An organisation needs the ability to revise these mechanisms while retaining the knowledge and methods its people rely on. Agree the export, handover and support arrangements that make that continuity practical.