Read more of this story at Slashdot.
Read more of this story at Slashdot.

AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same repositories, giving agents an opportunity to learn from previous work. Without that learning, an agent may repeatedly rediscover how to run tests, load tools it never uses, or spend dozens of turns revisiting the same error.Â
In our paper, TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents, we ask a simple question: can experience in a repository help an agent solve future tasks more efficiently? TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security, turns completed sessions into guidance for subsequent ones. It reduces waste without changing the agent or its underlying model, so developers can use it alongside the coding agents they already rely on.

We first profiled Claude Code on 60 tasks from SWE-Bench, a benchmark of real bug fixes in open-source Python projects, running each repository’s tasks in sequence. Three kinds of waste recurred:
We also tested three popular open-source plugins, each built around a fixed cost-saving strategy. Each lowered costs in some repositories but not others. Together, these findings suggest that effective cost control must adapt to the repository, rather than apply the same strategy everywhere.
Microsoft Research StorY
A three‑month pilot in a Midwestern bottling plant shows what happens when AI moves beyond chat and into decision-making, where constraints shift, stakes are real, and answers must hold.
After each session, TokenEconomizer reads the session’s log and updates what it knows about the repository. It then acts through three mechanisms, using only interfaces the agent already provides, as illustrated in Figure 2.
Micro-skills capture repository-specific knowledge in one-line notes. TokenEconomizer identifies recurring patterns across sessions and proposes corresponding notes. They are added to the agent’s starting context only after several sessions have confirmed the pattern, and retires the note if later evidence contradicts it.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Failure-loop intervention watches for an agent that keeps hitting the same failure. A judge model decides whether the agent is truly stuck or still making progress. If it is stuck, the agent gets a short note suggesting a different approach. In the open-source plugin, if the agent is still stuck after the note, TokenEconomizer alerts the developer rather than stopping the work itself.

The controller is designed to revise its decisions as new evidence arrives. Reported savings also account for the cost of TokenEconomizer’s own model calls.
We evaluated TokenEconomizer on 277 SWE-Bench tasks from six repositories, using five different AI coding agent and model combinations. For each repository, TokenEconomizer started with no prior session history.Â
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss. By contrast, lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation.

Savings grow as experience accumulates. Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task, shown inFigure 4. Each mechanism contributes savings on its own, and combining all three delivers the largest reduction.

The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
TokenEconomizer opens a path toward coding agents that learn not only how to solve tasks, but also how to spend computation more effectively. Its lightweight controller shows how experience can improve efficiency without changing the underlying agent or model.
The open-source plugin makes these directions accessible to developers and researchers. It is designed to run alongside existing AI coding agents without modifying their underlying models, requires no API key of its own, and keeps what it learns on the developer’s machine. We invite the community to use it as a starting point for exploring agents that become more resourceful with experience, adapting how they work, as well as what they know.Â
Opens in a new tabThe post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
“Kickers are too good now” sounds plausible when long field goals keep making the highlights. But in the November 9 episode of The Monday Morning Agent, our AI analyst found that long kicks went through at roughly the same rate across the seasons on file. Coaches were attempting them more often. We gave the agent written visualization standards so its charts would show both success rates and attempts per game, with the context readers need to judge the claim.
We started with two charts of league-wide field goal success rates. One began in 2024, with a vertical axis starting around 84%. The line climbed to 85.6%, and the improvement looked dramatic.
The other included every season from 2021 onward, started the vertical axis at zero, and showed the attempt count beneath each point. Success rates stayed around 84% to 86%. The recent rise looked much smaller in that context.
Changing the starting year and axis made the recent improvement look larger. Before your agent draws a trend, decide which choices it can make. Otherwise, it can produce a persuasive picture without giving readers the context to assess it.
The dictionary from episode two still applies. During a rehearsal, that earlier agent refused the kicking claim because its dictionary could describe a rush but couldn’t describe a kick. We expanded what it could define while keeping the requirement to use a saved definition.
This time, the play descriptions named field goals and included distances. We still had to decide where “long” starts. A recognizable play name doesn’t settle how to group it for analysis.
We examined distance bands separately. Kickers made roughly 95% of attempts under 40 yards. Success rates hovered around 80% for attempts from 40 to 49 yards and around 70% for those from 50 to 59 yards. Combining them into one rate can hide the differences the question asks about.
We introduced eight rules. Every trend includes all seasons on file, and axes start at zero. Each point shows its underlying count. Yellow identifies the subject of the claim, blue identifies the comparison, and gray marks the current season.
The chart title asks a question rather than announcing a conclusion. A source line gives the exact command that produced the data, so readers can rerun it. The chart shows only values the tool printed, without added trend lines or numbers typed by hand. Every chart uses the same layout, with success rates above attempts per game.
Writing these standards down gives the agent consistent instructions and gives readers a familiar way to compare results across episodes. Choose standards that fit your analysis, then specify how you’ll check each one.
A written standard needs a way to enforce it. We added a renderer, the program that turns the agent’s chart specification into a picture. It checks the chart specification against the rules before drawing the chart and refuses specifications that break them.
We deliberately requested a trend beginning in 2023. The renderer refused because the first season on file was 2021. Its response identified the rule instead of producing the shortened chart.
Give your agent a request that violates a standard and inspect the response. That check shows whether the restriction works when the agent tries to produce an output.
The longest kicks need particular care. The episode’s current-season sample contained only nine attempts in the group of kicks from 60 yards or farther. The next kick could noticeably change that rate. Displaying counts beside percentages helps readers distinguish a small sample from hundreds of attempts.
Showing success rates alongside attempts per game helped explain the ruling. In these plays, long kicks weren’t consistently becoming more successful; coaches were trying them more often. More attempts give fans more opportunities to see a long field goal go through.
We also corrected episode two’s results after finding an error in the tool. Our defined substitute for the tush push had a conversion advantage of 10.1 percentage points overall and 9.7 without Philadelphia, replacing the previously reported 8.3 and 7.8. Chart standards help readers inspect results, but we still have to check the calculations behind them.
On November 16, we’ll add a second agent to independently audit an audience-selected claim. We’ll examine where the agents disagree and how they investigate the evidence. The agent doing the analysis shouldn’t also grade its own work.
O’Reilly members can sign up on the O’Reilly learning platform to join us live on November 23, 8:00am–8:45am PT / 11:00am–11:45am ET. Everyone can follow along on Radar every other Monday, with a brief pause for the holidays. Watch episode three above, and we’ll build on it in the next session.
Imagine a CSS karaoke machine that displays a ball bouncing on text that incrementally highlights, based on progress through via speech recognition and running that kind of animation as a single predictable, holistic effect driven by one progress value.
Staggering Scroll-Driven Animations with animation-range-start originally handwritten and published with love on CSS-Tricks. You should really get the newsletter as well.
A second delivery of the same message can reveal more about an application than its happy-path demo. I like tools that make those cases easier to investigate while a change is still small. Last week’s launch of AWS Emulator brings part of that work onto your own machine: developers and coding agents can exercise AWS API interactions, inspect results, and try again without provisioning cloud resources for every attempt.
I wrote about the Emulator’s launch last week, and it’s my main highlight in this edition. There are also new ways to isolate workloads on shared infrastructure, give an agent narrowly scoped access to a backend, and prepare search applications for a regional disruption. Now, let’s get into this week’s AWS news…

AWS Emulator runs locally alongside your code and development tools. Image: AWS.
Headlines
Test your applications against AWS locally with AWS Emulator introduces a generally available local environment for AWS API interactions. Point the AWS CLI or AWS SDKs at its HTTPS endpoint, and your application keeps its normal SDK calls. The walkthrough shows a coding agent creating an Amazon DynamoDB table, writing an item, and reading it back. You don’t need an AWS account for emulated operations, and there is no additional service fee. The getting-started guide covers Linux packages, Docker for macOS and Windows, and agent-assisted setup; the source is on GitHub.
What I like is giving an agent concrete feedback while it works: run calls, inspect results, and revise the code. Matt Wood’s perspective on feedback for coding agents explores why that loop matters. Check the published fidelity results for your operations, because coverage varies and unimplemented calls return HTTP 501. The emulator doesn’t reproduce IAM and resource-policy enforcement or cross-account and cross-Region behavior, so test the complete application and its security configuration on AWS before production.
AWS Nitro Sandboxes bring hardware isolation to pods, tasks, and VMs. You can give an Amazon EKS pod or Amazon ECS task its own VM boundary while sharing an Amazon EC2 instance. That is useful for tenant workloads, build jobs, and agents executing generated code. Isolated mode separates workloads from their neighbors; confidential mode adds protection from node operators and cryptographic attestation. The Compute Blog deep dive explains both modes and their integrations. Nitro Sandboxes have no additional charge on supported instances; EC2 and applicable EKS Auto Mode or ECS Managed Instances fees still apply.
Last week’s launches
Here are some launches and updates from this past week that caught my attention:
Amazon Cognito supports OAuth 2.0 token exchange – You can now exchange a user’s token for one scoped to a downstream service in Amazon Cognito. The derived token can have narrower permissions and a single audience, with no refresh token; supplying an actor token also identifies the intermediary in an act claim. For an agent calling a backend on someone’s behalf, that makes delegation more explicit without forwarding the original token. The feature is available on Essentials and Plus plans in all Cognito Regions.
Claude Fable 5.2 is available on AWS – You can access the model through Amazon Bedrock or Claude Platform on AWS. Anthropic positions it as its most capable generally available model for technical work, including extended engineering tasks, scientific research, and quantitative finance. Bedrock has model-specific retention requirements: Fable 5.2 traffic is retained for up to 30 days for abuse detection, with a temporary zero-retention exception for eligible internal use.
Amazon ElastiCache Serverless adds search and aggregations – You can now combine vector, full-text, and hybrid search with server-side aggregations in Amazon ElastiCache Serverless for Valkey. That gives product discovery, recommendations, and agent-memory retrieval another option for querying frequently changing data. The capability requires Valkey 9, so create a compatible cache or upgrade an existing one. It is available in commercial Regions, AWS GovCloud (US), and the AWS European Sovereign Cloud.
Amazon OpenSearch Service adds managed cross-Region disaster recovery – You can now pair Amazon OpenSearch Service domains running version 3.7 or later in two Regions within one account. AWS specifies recovery objectives under one minute for data loss and under 20 minutes for recovery time. Planned switchovers preserve all writes; unplanned failovers can lose unreplicated data. The deep dive covers the remaining work: keeping configuration aligned and redirecting application writes after switching roles, plus secondary-domain and data-transfer charges.
Amazon CloudWatch Logs adds account-level retention policies – You can now set retention across all log groups or groups matching a name prefix in Amazon CloudWatch, with policies covering existing groups or only new ones. When policies overlap, the longest retention period applies. A platform team can set a baseline and retain security logs longer without updating every group individually. The feature has no additional charge and is available in commercial Regions except Middle East (Bahrain) and Middle East (UAE).
The next generation AgentCore runtime is now the default – New Amazon Bedrock AgentCore runtimes now use the next-generation runtime by default, so agent sessions start in under 2 seconds, scale up to 8 vCPU and 32 GB, scale to zero when idle, and bill up to 68% less memory, while each session still runs in its own isolated microVM.
For a full list of AWS announcements, be sure to keep an eye on the What’s New with AWS page.
Other AWS news
Here are some additional posts and resources that you might find interesting:
Optimizing cost per task and proving agent ROI on Amazon Bedrock – This guide follows an agent’s cost through repeated calls, reasoning, caching, and human review. I like its emphasis on measuring a completed task and comparing outcomes against a documented baseline. Its worked prices are illustrative; use your own workload and reconcile estimated costs with billing data.
How Zillow runs 100 million daily notifications on AWS Step Functions – Zillow describes reusable content modules and configuration-driven AWS Step Functions Express Workflows for more than 100 million daily notifications. The useful engineering details include idempotent modules, batching, campaign isolation, and testing expressions with the TestState API before deployment. Those patterns apply well beyond notification systems.
Upcoming AWS events
Check your calendar and sign up for upcoming AWS events:
AWS re:Invent – Join the community in Las Vegas from November 30 to December 4. Browse more than 2,200 sessions, including workshops and deep dives, or register for the free livestream to follow keynotes and selected sessions remotely.
AWS Community Days – Find community-led technical talks, workshops, and hands-on learning near you. Check each organizer’s page for current dates and registration details.
Join the AWS Builder Center to connect with builders, share solutions, and access content that supports your development. Browse here for upcoming AWS-led in-person and virtual events and developer-focused events.
That’s all for this week. Check back next Monday for another Weekly Roundup!
– Daniel Abib
This post is part of our Weekly Roundup series. Check back each week for a quick roundup of interesting news and announcements from AWS!