Engineering agents that create, review, and ship.
Aampe’s agents work across the product and the engineering team: generating campaign copy, carrying fixes through implementation and review, and learning from each customer’s behavior. Each workflow needs the right context, tools, evaluation, and human feedback.
The work
Agentic products and applied AI engineering workflows
Relay turns campaign intent into generated copy that marketers can shape and review. Patches takes frontend issues through clarification, implementation, live QA, and code review with the Aampe team.
Adaptive learning and real-time pipelines
The core learning pipeline moved from batch processing to real-time updates, connecting fresh features and behavioural signals to per-user decisions.
Relay + Prompt Shaper
From a marketer’s intent to useful, channel-aware copy.
Relay is Aampe’s LLM copy-generation engine. Prompt Shaper provides a conversational way to turn campaign intent into usable generation inputs.
Shape intent before generation
Prompt Shaper turns a marketer’s intent into the input Relay needs. Brand context, campaign inputs, and context profiles help the model produce copy suited to the channel and audience.
Evaluate within the product experience
Scoring dimensions, live approval, and override flows connect LLM evaluation to the person reviewing the generated copy. Editable modifiers steer tone, length, and format, so refinement stays inside the product experience.
In practice
Reported first-pass quality improved from 70–75% before Prompt Shaper to about 95% afterwards.
Patches: an always-on engineering agent
Give small frontend fixes a path through the backlog.
Small bugs and rough edges in Aampe Composer kept losing priority to larger features. The overhead of picking up each ticket made even simple fixes expensive. Patches lets designers and customer-facing teams move those fixes forward, while engineers retain final code review.
From intake to a reviewed fix
A labelled Linear ticket starts the loop. Patches asks clarifying questions, plans the solution, implements it, and opens a draft GitLab merge request. A live frontend preview lets the reporter test the change and request revisions. After reporter sign-off, a developer reviews the code; approval allows Patches to merge and the deployment pipeline to ship.
Consolidate three agents into one
The first version split triage, implementation, and review across three agents. Handoffs consumed tokens and lost context. One agent carrying the whole job reduced that coordination overhead and kept the context together.
Check for work before calling the model
The initial agent searched for new work itself. A deterministic harness replaced that polling, checking for new tickets and changed merge requests every minute and invoking the agent only when something needed attention.
Make the work inspectable
The frontend harness runs against a staging backend and posts screenshots into Linear. A dashboard exposes ticket progress, session logs, tool calls, and cost.
In practice
In the month covered by Aampe’s 2 June 2026 report, Patches closed 99 issues for around $625. That spend also included work on 43 unfinished issues. Reporters could advance and QA fixes while engineers focused on the final code review.
Read Aampe’s Patches engineering accountAdaptive learning for each customer
An individual learning loop across messages, timing, channels, and product experiences.
The learning pipeline connects fresh behavioural signals to per-user agents, combining online updates with adaptive decisions across customer interactions.
Balance exploration with what is already working
Thompson Sampling is one of the algorithms Aampe uses to balance trying uncertain options with applying what an agent has already learned. The explore-exploit balance adapts for each person.
Keep evidence responsive to changing behavior
The realtime pipeline supplies recent signals. Decay reduces the influence of old evidence as context changes, while feedback is shared across messaging and in-app experiences.
Think beyond a single reward
Clicks, conversion, and longer-term engagement answer different business questions. Reward design must consider attribution and multiple business goals across combinations of content, timing, and channel.
In practice
The move from batch processing to real-time updates allows agents to act on recent behaviour. The linked article explains how adaptive learning balances exploration with the evidence already available.
Read the Thompson Sampling articleAn educational game illustrating exploration and adaptation.
Results & progress
Learning, creating, and shipping with agents.
Marketers can shape and review generated copy in Relay. Issue reporters can validate frontend fixes in Patches before engineering review. Per-user learning agents can make decisions with fresh behavioural signals.
Background on the learning pipeline, real-time migration, and Applied AI work.
Applied AI capability mapping
- Product & LLMs
- Agentic AI product and workflow design
- LLM integration
- Prompt engineering
- Context engineering
- Agent engineering
- Agent harness design and deployment
- Tool calling and deterministic API integration
- Evaluation & cost
- Production feedback and continuous evaluation loops
- LLM cost attribution
- Learning systems
- Recommendation and personalisation systems
- Per-user learning agents and real-time decisioning
- Batch-to-real-time learning pipeline migration
- Online feature and behavioural signal pipelines