daily issue · September 18, 2026
How did DoorDash use agents to clean up feature flags?
An agent can propose a change or return a structured decision. Keep the acceptance rules and the authority to approve with the team.
Direct answer
DoorDash built a multi-agent workflow to identify and remove old feature flags across its repositories, according to InfoQ. Its evaluation measured proposals that turned into usable pull requests. For a comparable pilot, start with a flag whose rollout is already understood, verify that removing it leaves intended behavior intact, and require a code owner to approve the result.
Edited by Joe Cervino, Founder and Editor
Published
DoorDash’s case gives a concrete task for evaluating automation. Select flags that are safe to remove, check the behavior before and after the change, and send questionable patches to a code owner.
Thesis movement
Industry Matrix
Full public entity map from current Pulse signals versus the prior equivalent window.
Agentforce
Compare the operator’s claim with reviewed outcomes for one allowed cohort of existing warm leads.
Open the evidence from @jasonlk on X- Movement
- +27 proof
- Evidence
- 0 → 27
- Actionability
- 62 → 78
Evidence strengthened into Act Now on 1 signal across 1 source. SaaStr attributed part of its outreach work to an Agentforce sales agent.
Opinion. What did DoorDash’s feature-flag workflow establish?
DoorDash’s case gives a concrete task for evaluating automation. Select flags that are safe to remove, check the behavior before and after the change, and send questionable patches to a code owner.
TypeSafe’s paid decision-model access illustrates another way to limit a delegation. A team can define the valid output and an escalation point before asking an agent to propose any cleanup.
TypeSafe reports paid access to Jev’s structured decision model
TypeSafe published paid input pricing and free output for Jev’s structured choices, scores and probabilities. A comparison with previously reviewed labels can show which outputs still need attention.
Sources: @shannholmberg on X
The Sweep. Which changes matter when an agent crosses a system boundary?
DoorDash’s deployment, supplier changes and the legal product launch bring different operating boundaries into focus. For each workflow, check who can review a change and which data crosses into another system.
DoorDash uses multiple agents to remove old feature flags
Review which flag can be removed, then require a test and code owner review before merging the cleanup.
Sources: InfoQ
An operator reports repeated provider switches for recurring workloads
Check how much task context and review history you can export before committing to one supplier.
Sources: @JayaGup10 on X
OpenAI packages GPT-6 Astra for legal workflows
Test one legal workflow with approved material and confirm which plugin can read or write case data.
Sources: @cgtwts on X
Discourse and OpenAI fixed reported single sign-on weaknesses
Review the vendor patches and verify your approved identity configuration before restoring sensitive access.
Sources: @sterlingcrispin on X
AI Industry News. Which new integrations change an operating decision?
Model hosting and compressed checkpoints can reduce cost under some workloads. Review each supplier’s data-sharing terms and test the same completed task before committing to a route.
The issue spans payment workflows, deployment and governance. Ask which new capability is independently testable and who assumes responsibility for an unexpected change.
The selected demos show code and media workflows using different tools. Measure the completed result and preserve a way for an operator to stop or rerun the work.
Preparation for a resumed task includes instructions, memory and context. Validate what the agent recovers after a failure and whether the recovered procedure is current.
The audit findings and deployment timings point to separate checks on evidence and operations. Test the score on the work you own and time the steps users experience.
Models, Routing & Open Source
PrismML compares compressed Ternary Bonsai with its full-precision base
Track peak memory and score regressions on your own tasks before choosing the smaller checkpoint.
Sources: @PrismML on X
OmniRoute reports a lower DeepSeek price through another provider
Measure effective cost and quality using equivalent work before switching traffic based on the posted rate.
Sources: OmniRoute by Cheaper Inference | #1 Open Source AI Router
AgentSky opens a marketplace for coding agents and handoffs
Test a small handoff and inspect which parts of the task context the next provider receives.
Sources: @quxiaoyin on X
AWS adds Moonshot AI’s Kimi K3 to Bedrock
Check how prompt caching and access controls affect a representative long-document task.
Sources: AWS ML Blog
Bolt offers more coding usage when developers share sessions
Review exactly what training data would leave your workspace before taking the extra usage.
Sources: @EvanKirstel on X
GitLab adds hosted open models to Duo Agent Platform
Compare credits for the same completed coding task before picking an open model for routine work.
Sources: @bstaples on X
Dax describes the cost of absorbing DeepSeek launch demand
Ask a capacity provider how it handles launch spikes and when a workload can move to another host.
Sources: @thdxr on X
Industry moves
A consumer report finds more Americans paying directly for AI
Test willingness to pay in the group your product actually serves. Adoption in one age group need not transfer to another.
Sources: @a16z on X
Squad reports reaching a new recurring revenue milestone
Compare the company’s reported revenue growth with customer retention before using it as an adoption benchmark.
Sources: @tibo_maker on X
LlamaIndex shows why tabular answers need their original context
Give a test agent table data with shared labels and check that its answer preserves units and hierarchy.
Sources: @llama_index on X
PayPal reports lower inference use in a WebMCP checkout test
Who for: Checkout teams comparing machine-readable tools with browser navigation.
Try one consented checkout and record average effort along with which actions still require approval.
Sources: @sarah_edo on X
Figure reports high throughput for its robotics data pipeline
Ask which types of consented data reach the training workflow and how unusable recordings are screened.
Sources: @adcock_brett on X
Snowflake integrates with STACKIT for German sovereign workloads
Verify residency and access terms for the data and workloads you would put on the combined service.
Sources: Snowflake AI
A software author proposes a deterministic core for agent workflows
Decide which steps must run the same way on every request before giving an agent discretion elsewhere.
Sources: @DavidKPiano on X
Palantir’s CEO warns of customer intellectual property migration risk
Ask a prospective provider where your company’s data goes and what you can retrieve at the end of the contract.
Sources: @WallStRollup on X
Siemens reports AI controlling dosing at a Belgian brewery
Review which operating conditions still require staff supervision before granting a similar system control.
Sources: @EvanKirstel on X
AgentCloak replaces sensitive values locally before sending prompts
Test a non-sensitive example and see whether the substitutions preserve meaning and restore the expected fields.
Sources: @rohanpaul_ai on X
Harness, Skills & Tools
A Codex demo integrates two generated three-dimensional assets
Test the same design handoff on an approved asset and have someone check the resulting file before reuse.
Sources: @MengTo on X
Mollick estimates token cost for a long multi-agent media task
Estimate the full budget of an exploratory run before using a one-off result to price recurring work.
Sources: @emollick on X
A robotics demo delegates high-level decisions to GLM
Use a simulator and fixed safety rules to check whether the tool calls remain appropriate across repeated tests.
Sources: sentdex
Knowledge, Context & Prompting
OpenAI documents unexpected behavior after summarizing agent context
Check whether safety and task instructions remain intact when an agent resumes from a shortened context.
Sources: Simon Willison
Claude Code adds optional support for AGENTS.md
Compare which repository instructions are loaded with and without a legacy file before sharing one ruleset across tools.
Sources: @trq212 on X
Memorable saves successful agent procedures for later reuse
Run a repeated task after updating the source process and check whether a stored procedure still applies.
Sources: @garrytan on X
Theo argues compaction should preserve a coherent task history
Test a long task with an early constraint and see whether it survives a summarized handoff.
Sources: @theo on X
Evaluation, Security & Ops
Vercel reports faster steps in its global deployment system
Time artifact publication and domain readiness separately in your own rollout.
Sources: @rauchg on X
AI Employees. Which work needs a person to approve the outcome?
The retained case studies show recurring work in support, advertising and coding. Set a named owner for oversight and compare the amount of reviewed work with the original process.
NVIDIA tests Git as shared memory for research agents
Observe how test agents record decisions and resolve a conflict before making shared memory persistent.
Sources: @mark_k on X
SaaStr reports revenue improvement while operating production AI agents
Ask which workflows contributed to the result and what staff still reviewed. The reported outcome covers several agents together.
Sources: @jasonlk on X
Kulina reports increasing ad campaign volume with an assistant in Slack
Compare how operators approve spend and review changes before expanding campaign volume.
Sources: @aakashgupta on X
Microsoft is deploying agents for customer support and staff assistance
Compare routing and frontline guidance with the escalations that still need a person.
Sources: @rohanpaul_ai on X
An operator reports deploying a Google Ads agent under lead pricing
Choose a capped campaign and specify who can authorize changes to its daily budget.
Sources: @codyschneider on X
A bank gives engineering teams dedicated inference for coding agents
Test whether the engineers can choose, monitor and roll back their own model deployment under the new arrangement.
Sources: Together AI Blog
SpaceXAI demonstrates Grok Bot running a marketing workflow
Ask a human to review a test campaign’s creatives, destinations and proposed spend before allowing it to launch.
Sources: @joshkim on X
Kody connects external secret providers without showing agents the secret
Revoke a disposable credential during a test and confirm which actions stop.
Sources: @kentcdodds on X
Meta describes app controls in Muse for Mac
Check consent and confirmation requirements for each app before asking the assistant to work across accounts.
Sources: @alexandr_wang on X
An operator reports subscriptions from an AI-managed ad campaign
Validate which signups convert under a capped campaign and compare them with an independently tracked baseline.
Sources: @codyschneider on X
Claude Projects beta can continue parallel work in the cloud
Try a reversible task, close the laptop and check who can inspect the cloud work and approve completed changes.
Sources: @T_Zahil on X
Anthropic reports more model research tasks where Claude takes the lead
Distinguish its reported task delegation from verified end-to-end outcomes and review where people still intervene.
Sources: @kimmonismus on X
Startup Highlights. Which funded software announcements qualified?
The funded products span data protection, loan operations, construction finance and agent monitoring. Choose a recurring exception to walk through before buying so the team can see what the product handles and what remains with a person.
Mind raises $72 million for AI data protection
Who for: Security administrators reviewing what data employees can share with AI tools.
The Series B supports software that helps companies identify sensitive data and control where AI tools can send it. Crosspoint Capital Partners led the round.
funded · ai SaaS
Sources: siliconangle.com
Adaptive raises $30 million for construction accounting AI
Who for: Construction finance teams reviewing how accounting records reach project staff.
The Series B backs software for accounting workflows used by construction companies. Tidemark led the financing.
funded · ai SaaS
Sources: devcuration.com
Kastle raises $24 million for AI loan operations
Who for: Loan operations teams evaluating borrower handoffs and exception review.
Insight Partners led a Series A for the loan operations software provider. Buyers can inspect which decisions and borrower records the system handles.
funded · ai SaaS
Sources: devcuration.com · founderland.ai
Lightspeed backs Raindrop’s AI agent monitoring platform
Who for: Reliability teams operating agents that can fail without an application error.
Lightspeed says it led Raindrop’s Series A, taking total funding to $50 million. The company monitors autonomous agents for production anomalies.
funded · ai SaaS
Sources: lsvp.com
Octa raises $3.5 million for AI accounting workflows
Who for: Accounting teams deciding where workload automation still needs a review queue.
The seed round supports Octa’s software for tasks used by accounting firms. MEVP led the financing.
funded · ai SaaS
Sources: techscoop.io
Zero raises a $10.3 million seed for revenue operations software
Who for: Sales operations teams combining prospect data with existing pipeline records.
The Helsinki vendor offers software that handles parts of prospecting and revenue operations across company systems.
funded · ai SaaS
Sources: b2bdaily.com
EnforceShield raises €1.7 million for copyright enforcement software
Who for: Rights teams reviewing copyright claims.
Test evidence review and signoff on a sample of suspected violations before evaluating automated enforcement.
funded · ai SaaS
Sources: foundernews.eu
Resources. Which checks can a team run on a known failure?
The benchmarks, SDK changes and runtime notes help teams measure different parts of an agent workflow. Pick an observed failure, then choose a resource that can reproduce the conditions and expose which step failed.
DeepSeek Harness makes model and execution components replaceable
Swap one component at a time on a fixed exercise and check whether the harness still respects the task boundary.
Sources: @Freyabuilds on X
OpenAI publishes a framework for reporting model misalignment
Give an evaluator a known incident and check whether the outlined reporting process identifies an owner and review record.
Sources: InfoQ
Anthropic’s skills guide describes how instructions declare their triggers
Give a test skill two similar inputs and see whether it activates only for the intended job.
Sources: @godofprompt on X
EvoSkill turns agent failures into reusable procedures
Check that a saved procedure improves the next similar task and is not recalled for an unrelated one.
Sources: @rohanpaul_ai on X
AWS documents a healthcare agent migration to Bedrock AgentCore
Use synthetic records to compare existing retrieval and approval behavior before migrating.
Sources: AWS ML Blog
WSO2 releases an open control plane for agent governance
Inventory a test set of agent identities and check who can review or change each allowed action.
Sources: InfoQ
A Together AI manager keeps feature decisions with people
Assign a human owner to approve the boundary of one feature before agents implement it.
Sources: @aakashgupta on X
Financial Datasets adds a Claude plugin for public-company metrics
Compare one returned metric with the underlying company disclosure before relying on it in a report.
Sources: @findatasets on X
Trylle gives coding agents isolated cloud workspaces for pull requests
Verify the task, tests and reviewer on a disposable issue before allowing direct repository changes.
Sources: @catalinmpit on X
Univer Office Harness adds review and isolated work for office tasks
Try a document update and verify that the reviewer can inspect its source changes before approval.
Sources: @LeonUniver on X
Anthropic proposes external measures for assessing AI research and oversight
Ask which evidence an external reviewer could inspect when judging delegation claims and oversight practices.
Sources: @AnthropicAI on X
OpenAI Agents SDK aligns conditional approvals with validated arguments
The release notes describe improved handling for conditional approvals and resumed sessions. Confirm that the approval prompt matches the arguments the tool receives.
Sources: newsopenai.net
Claude Agent SDK for Python updates its bundled CLI
The reported update changes the CLI packaged with the Python SDK. Check which CLI behavior your integration relies on before upgrading.
Sources: llm-digest.com
llama.cpp adjusts how embedded model weights are loaded
The update reports changes to alignment and file-pointer handling for embedded model weights and adapters. Verify behavior on a package you ship.
Sources: aiaiaiaiaiaiaiaiaiaiaiaiaiaiaiai.com
Flyweight publishes an open runtime for mixture-of-experts models
The initial release provides a native inference runtime for supported GGUF models. Check compatibility with the model you already use for a local workload.
Sources: myprivateclaw.com
Vals publishes a benchmark for sustained changes to applications
Vibe Code Bench begins with a passing app and checks sequences of dependent requests. One incomplete change stops the run, revealing failures that a one-turn score may hide.
Sources: aicoder.com
Epoch AI launches reviews of widely used benchmarks
The initiative makes underlying benchmark checks visible and records whether reviewers found them sound, flawed or insufficiently documented.
Sources: bsky.app
MLPerf adds tests for retrieval and agent inference
The report describes additional workloads for retrieval and agents. Select the benchmark whose request pattern resembles the service you operate.
Sources: allatra.media
Zing 0.5 reports text and keyboard controls for world models
The article describes conditioning one generative model on instructions and keyboard input. Compare whether the simulation preserves state across interactions.
Sources: cctest.ai
Meta researchers test a separate memory agent for another agent
The study describes a helper that maintains reminders for a main agent. Test whether the reminder actually changes what the task agent carries into its next action.
Sources: deeplearning.ai
Google’s Android Bench adds multi-day engineering tasks
Coverage describes a shift toward longer tasks with checks on the completed engineering work. Use a comparable task when evaluating an Android coding workflow.
Sources: androidcentral.com
An Oracle benchmark measures latency as persistent agents scale
The published account describes read-only agent requests and reports how latency changes at higher concurrency. Check tail latency under your own request pattern.
Sources: superpowerdaily.com
Claude Code adds request IDs to assistant transcript entries
The changelog describes request IDs and structured metadata in transcripts. Verify that operational records can follow a task across its related events.
Sources: changelogs.core-directive.com
Sanity adds validation and authoring support to workflows
Who for: Teams replacing an existing tool for the specific workflow described.
The changelog includes workflow definition checks and changes to asset uploads. Test a staged review before assigning the new workflow to a content team.
Sources: sanity.io
Gemini Managed Agents adds file and credential interfaces
The account highlights changes to file handling and permission-related APIs. Test which actions can finish with the minimum access needed for one approved workflow.
Sources: traeai.com
Posit updates commons, shinychat and ellmer for data agents
Who for: R and Python teams testing self-service analysis.
Give a data agent a vetted calculation and check whether the result is marked verified and cited.
Sources: opensource.posit.co
Microsoft’s Frontier Playbook sets three levels of AI autonomy
Who for: Operations leaders setting approval boundaries.
Pick one workflow, choose how much the agent may do, and name the person who stays accountable for the result.
Sources: mspbusiness.com
A WebMCP test compares structured tooling with browser-based tasks
Who for: Teams testing how agents interact with websites.
The report compares a Jev and Mercury setup with Astra using two browser methods. Run the same completed-task check for each setup before using the cost comparison.
Sources: @0xidanlevin on X