Fresh context for agents. Rebuilt every 12 hours. Provenance by Synorb.
Updated 2026-11-19 00:00 UTCnext update 2026-11-19 12:00 UTC
Confirmed · 4 Manifests838 words

Coding agents finally get budgets, approvals and a bill tied to the pull request

LangSmith now ties agent spend to pull requests, Snowflake cut its AI bill with per-user quotas, Qodo automates routine approvals, and one firm lets agents watch for work without waking the model at all.

Key facts
  • LangSmith now links pull requests and merge requests to agent threads and lets teams compare cost per pull request before and after a model change. Self-reported m:…1115
  • Snowflake says per-user quotas in its Cortex AI Gateway cut AI costs by 16%, after an engineer's internal Slack bots spent $1,100 in a single day through automatic model selection. Self-reported m:…9662
  • Qodo's Auto Approval Workflows, in beta, automate pull request approvals based on findings by severity and category, and withdraw approval if a new commit means the change no longer qualifies. Self-reported m:…4973
  • Scale Venture Partners' program watches run scheduled Python code against read-only tools, and a watch that finds nothing runs no model at all. Self-reported m:…8010

The bill arrives per pull request

Coding agents have been easy to start and hard to account for, because the spend showed up as one large token bill with no connection to the work it produced. LangSmith now ties that spend to the unit engineering teams actually manage: the pull request. Its update for coding agents links GitHub pull requests and GitLab merge requests to agent threads, and provides a dashboard to track spend by pull request state, including open, closed, and merged. Enterprise AI application spend on coding reached $34 billion in 2026, according to LangSmith, which is the scale of the question teams are now asking: what did each merged change cost, and was it worth it. m:…1115

The tool is aimed at the two kinds of waste that are hardest to see. First, the platform lets teams compare cost-per-pull-request metrics before and after model changes, so a switch to a stronger model can be judged on what it cost per merged change rather than on a benchmark score. Second, it tracks unattributed spend for sessions that do not result in a pull request, covering code review and investigation work that otherwise disappears from the accounting. Threads are labeled by work type, such as features, bug fixes, investigation, and code review. Long threads concentrate spend because they carry the entire conversation context, and clearing context between tasks is recommended to reduce agent spend. The feature is currently in public beta on LangSmith Cloud, and it supports Claude Code, OpenAI Codex, Cursor, GitHub Copilot, and Agents. m:…1115

Quotas that bite within minutes

Snowflake's answer to the same problem is a hard limit rather than a dashboard. The company implemented per-user quotas in its Cortex AI Gateway to control AI spending, moving from team-wide budgets to four standardized tiers, and says it achieved a 16% reduction in AI costs using quotas. The trigger was concrete: an engineer's internal Slack bots spent $1,100 in a single day due to automatic model selection. Per-user quotas block AI requests within minutes of reaching a limit, which turns a runaway loop into a short incident instead of a month-end surprise. m:…9662

The mechanics are deliberately boring, which is the point. Quota membership is driven by a user tag, and Snowflake built a director-only application on its App Runtime to manage quotas, so engineering directors control individual limits directly. An account-wide safety net quota covers users without a specific tier, and daily limits provide better protection against runaway loops than monthly limits alone. Snowflake piloted the quota system in monitor-only mode before enabling enforcement, and the highest tier requires a written business justification. The lesson for other teams is that the control that works is the one set before the spend happens, at the level of the individual user, with a daily ceiling. m:…9662

Approvals, and work that never wakes the model

If spend is one half of the control problem, review is the other, and Qodo is automating the routine part of it. The company introduced Auto Approval Workflows in beta, which automate pull request approvals based on findings by severity and category. Approvals can be limited to changes with a small footprint, and guards prevent automatic approval for guarded repositories or paths. Approved changes can be marked as approved by the company's bot or given a label, and Qodo withdraws its approval if a pull request no longer qualifies after a new commit. Alongside it, Qodo introduced Agent Session Review in research preview, which gathers every pull request created in a session across repositories, highlights deviations from the plan and implementation choices, and provides a verdict on risks tied to evidence. m:…4973

Scale Venture Partners takes a different route to the same goal: do not wake the model unless there is something to do. The firm built an in-house agent platform called Deckhand, which runs background coding agents and personal agents, and added program watches, short Python programs that run on a schedule against read-only tools. The watches use Monty, a minimal Python interpreter for running model-written code in a sandbox, and CodeMode, which allows models to compose tool calls as code. A decision model judges program outputs to determine if an agent should be woken, and a watch that finds nothing runs no model at all. Watches run at most every ten minutes, the worker checks that every tool is marked read-only before execution, Monty runs code in a pool of subprocess workers to isolate crashes, and a watch stops after three consecutive failures. The implication is that the cheapest agent run is the one that never happens, and the infrastructure to decide that is now part of the product. m:…8010

Taken together, these four moves describe a shift in how coding agents are managed. The first wave was about capability: can the agent write the code. This wave is about governance: what did it cost, who approved it, and was the model even needed. The teams that adopt these controls early get the numbers to defend the next expansion, and the teams that do not will keep discovering the cost of their agents the same way Snowflake did, one surprising day at a time. m:…1115 m:…9662 m:…4973 m:…8010

What is still unresolved

  • Whether cost per pull request becomes the standard way teams compare coding agents, or whether vendors settle on different metrics.
  • How often automated approvals let through changes a human reviewer would have caught, and how teams will measure that.
  • Whether program watches and similar no-model monitoring spread beyond the firms building their own agent platforms.

Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.

Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.

Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

  1. Get a Synorb Key. Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool synorb-get-demo-key (no header needed), or REST POST https://api.synorb.com/api/connect/demo-key with a JSON body such as {"client":"your-agent-name"}. The key is the api_key field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
  2. Read each cited Manifest by id. REST: GET https://api.synorb.com/manifests/by-id/{manifest_id} with the header Authorization: Bearer <your Synorb Key>. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
  3. Check the claim. Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.
Cited Manifest ids (4)
  • 1795014732631041115
  • 1795041445893289662
  • 1795007210172004973
  • 1795036667348248010
for id in 1795014732631041115 1795041445893289662 1795007210172004973 1795036667348248010; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done

Provenance ledger

ManifestTitleSourceClaims usedPublished
1795014732631041115LangSmith for Coding Agents: Now track spend by outcomelangchain-blog1795015066577042645, 1795015066589042647, 1795015066593042655, 1795015066569042642, 1795015066591042650, 1795015066610042659, 1795015066627042662, 1795015066633042663, 179501506664404266711-18 00:00
1795041445893289662How Snowflake Uses Cortex AI Gateway to Control Runaway AI Spendsnowflake-engineering-blog1795041663818291783, 1795041663824291784, 1795041663830291785, 1795041663837291786, 1795041663843291787, 1795041663850291788, 1795041663857291789, 1795041663870291791, 1795041663876291792, 1795041663882291793, 179504166388929179411-18 00:00
1795007210172004973Introducing Auto Approval Workflows and Agent Session Reviewqodo-publishing1795007554468005900, 1795007554471005902, 1795007554475005906, 1795007554472005903, 1795007554474005905, 179500755447600590711-18 00:00
1795036667348248010Letting our agents write their own watchers with Monty and CodeModescale-ventures-blog1795036912886249564, 1795036912890249568, 1795036912887249565, 1795036912882249561, 1795036912884249562, 1795036912888249566, 1795036912889249567, 1795036912891249569, 1795036912892249570, 1795036912892249571, 179503691289324957211-18 00:00

Related stories