Cleric Starts Read-Only and Sends AI SRE Fixes as Pull Requests

Cleric is the default AI SRE agent for mixed stacks because its price, scopes and fix path are published. Bits AI is cheaper if Datadog is your only backend; Traversal and Resolve suit buyers who need deployment control.

By Rajesh Beri·November 19, 2026·14 min read
Share:
A pager and a laptop on a dark desk at night, the laptop showing a wall of red alert graphs, with an empty office chair pushed back as if the on-call engineer just stepped away.

Illustration generated using AI

For most platform teams on a mixed observability stack, buy Cleric first: it publishes a per-investigation price, starts read-only, and delivers fixes as pull requests you approve. If your telemetry already lives in Datadog, Bits AI SRE is the cheapest investigation on the market at about 4 credits each, and the price of that is lock-in. Traversal is the pick for a large enterprise that needs on-premises or bring-your-own-cloud deployment and will negotiate a contract. Resolve AI publishes its rates but bills by token, so you cannot know an investigation's cost until you run one. PagerDuty's SRE Agent (Paige) loses this comparison: it is a responder's assistant metered per question, and its autonomous mode is still listed as "coming soon."

Prices below were checked on each vendor's live pricing page on November 19, 2026. The workload used to compare them is 400 automatic investigations a month (roughly 13 alerts a day worth a look), with read-only access to metrics, logs, traces, deploy history and one GitHub org.

Agent Best for Default access Can it change production? Published price (checked 2026-11-19) ~Cost at 400 investigations/month Do not buy if
Cleric Mixed stacks, teams that want a known price Read-only Fix PRs and ticket updates, only when you allow it $0.50/credit; investigation = 10 credits ($5); Pro $1,500/mo, Scale $3,000/mo $3,000/mo (Scale; Pro's 3,000 credits cover 300) You need on-prem or your own model
Datadog Bits AI SRE Teams already standardised on Datadog Read-only investigation Yes, inside admin guardrails (Preview) AI Credits: $500 per 500 credits/mo annual, $1.30/credit on-demand; ~4 credits per investigation ~$2,000/mo annual (four 500-credit blocks) Half your telemetry lives outside Datadog
Traversal Large enterprises, regulated data "Read-only access, always" "Self-healing" offered; permission model not published Contact sales Unknown You need a price before a demo
Resolve AI Engineering-heavy orgs that want VPC or single-tenant "No write access" Claims mitigation; terms unclear $4 input / $20 output per 1M tokens (Pro); 1 Resolve Unit = $1 prepaid Unknown until you run one; ~$2,000/mo if a run costs $5 (our assumption) Finance needs a fixed monthly number
PagerDuty SRE Agent (Paige) Teams already paying for the PagerDuty platform Reads incidents, runbooks, logs Human-approved runbook actions; autonomous mode "coming soon" 4 AI Actions per request; Essential $12,000/yr includes 5,000 AI Actions Plus ($48,000/yr) covers ~333 requests/mo You want an agent that investigates on its own today

Which AI SRE Agent Should You Buy?

Buy the one whose access model your security team will sign off on this quarter, then prove its accuracy on your own past incidents before you pay for a year. An AI SRE agent is software that takes an alert, queries your telemetry and code, and returns a root-cause hypothesis with evidence, and some can then propose or apply a fix. Every vendor here claims high accuracy on its own customers. The one independent benchmark worth reading says frontier models are much worse at this than the case studies imply.

That benchmark is ORCA-bench, 1,079 root-cause tasks built from six days of metrics, logs and traces on the OpenTelemetry demo shop. The best model reached 25.3% root-cause accuracy on medium tasks and 10.0% on hard ones, and hallucinated root causes in 7.2% to 40.2% of incident reports depending on the model. Read it with one caveat: Traversal researchers are co-authors, and a vendor has an interest in showing that raw models fail without its harness. The direction still matters. Vendor accuracy numbers come from incidents the vendor chose, on stacks it tuned for. Your incidents are the only test that predicts your result.

Buyers also have little appetite for unchecked automation. In a 696-person survey reported by The Next Platform (sponsored by NeuBird, an AI SRE vendor not compared here), 60% named trust as the blocker, 59% wanted near-perfect accuracy before adopting, and only 8% had AIOps in production.

What Each Agent Is Allowed to Touch in Production

The access model is the decision that is hardest to reverse, so start there. All five start read-only. They differ on what happens when you want more.

Cleric is read-only by default, with "write access when you're ready," and fixes arrive as pull requests for your approval. Its integration list spells out the scopes: a read-only IAM role on AWS, Kubernetes RBAC on EKS, and PR creation on GitHub or Bitbucket only when you enable it. It can post updates to incident.io and file Linear issues. This is the clearest permission model of the five, and you can hand it to a security reviewer as written.

Datadog Bits AI SRE separates investigation from action. Bits Remediation can run an infrastructure action "only when a guardrail covers the action and the resource it targets," and admins decide who must approve. The actions are real: rolling back a Kubernetes deployment, rebooting an EC2 instance, invoking a Lambda. One-click infrastructure actions are in Preview and are not supported on Datadog's government sites. Bits Code can open pull requests for code-related root causes.

PagerDuty's Paige can run runbook actions such as service restarts and Kubernetes remediation with human approval, and you set the rules for what needs approval. Autonomous resolution is listed as coming soon. Its documentation describes it reading incident details, looking up service dependencies and adding notes. Each workflow integration needs "Allow Paige access" ticked.

Traversal says "Read-only access, always" on its security page, while its home page sells a "Self-healing" capability that turns diagnoses into automated fixes. Neither page explains how the write path is granted or scoped. Ask for that document before procurement, not after.

Resolve AI states "No write access: Your data remains unchanged", while its home page says it "mitigates and restores service in minutes, not hours." Mitigation without write access means a human or another tool does the work. Get the vendor to say which in writing, because the answer determines whether you are buying an investigator or an operator.

AI SRE Pricing: Credits, Tokens and Platform Bundles

Three billing models are in use, and each one grows with alert volume in a different way. This is the line item that surprises buyers in month four.

Per-investigation credits: Cleric and Datadog

Cleric's pricing is the most legible: credits cost $0.50, an investigation is 10 credits ($5), verifying one production change is 2 credits ($1), and chat is 1 credit per started minute after the first 30 seconds. The plans are Team at $700/month (1,400 credits), Pro at $1,500 (3,000) and Scale at $3,000 (6,000). At 400 investigations a month you need 4,000 credits, which puts you on Scale. The detail that matters in an alert storm: repeat alerts that join an open Issue, retries and follow-up passes are free. If 400 alerts collapse into 150 Issues, you pay for 150.

Datadog prices Bits through AI Credits at $500 per 500 credits per month on an annual commitment, or $1.30 per credit on-demand. Bits Investigation is listed at about 4 credits, with the caveat that use "may vary with task complexity." That makes 400 investigations about 1,600 credits: $2,000/month if you commit to four 500-credit blocks, or about $2,080 on-demand. Datadog says its average cost per investigation fell from $25 at launch to about $4 by October 2026. Remember that the AI Credits sit on top of the host, APM and log bill that makes Bits useful in the first place.

Per-token: Resolve AI

Resolve AI publishes token rates: $4 per million input tokens and $20 per million output on its Pro models, $12 and $60 on Pro Fast, with 1 Resolve Unit equal to $1 of prepaid usage. Investigations, chats and scheduled tasks all bill by tokens consumed. There is no per-investigation figure. If a run reads 1M input tokens and writes 50K output tokens, it costs $5 (that token count is our illustration, not a Resolve figure), and 400 such runs cost $2,000. Resolve does provide per-run spend limits, a monthly org cap and live cost per investigation (its feature table lists per-team caps only on Enterprise), so set the caps on day one. The Standard plan is pay-as-you-go by card, capped at 100 users and 5 integrations. VPC and single-tenant deployment, customer-managed keys and bring-your-own-model sit on the Custom plan, which has no published price.

Platform bundle: PagerDuty

PagerDuty's reliability platform charges an annual platform fee: Starter $2,800, Essential $12,000, Plus $48,000, Ultimate custom. Those include 1,000, 5,000, 16,000 and 160,000 AI Actions respectively. Paige consumes four AI Actions per request or nudge click, so Essential covers about 1,250 Paige requests a year and Plus about 4,000. Extra AI Actions come in credit packs priced through sales, and credits expire annually. PagerDuty is offering additional AI Actions on Essential and above with no credit charge for the first year to customers who sign by January 15, 2027. That makes the first-year bill look cheaper than the renewal will be.

No price: Traversal

Traversal lists no price, only a "Get Pricing" form. Its named customers (PepsiCo, DigitalOcean, Cloudways, American Express) suggest enterprise contracts, so expect an annual commitment and plan for a sales cycle.


Integration Footprint: Which Credentials You Hand Over

Each agent needs credentials for every system it reads, and every credential is one your security team has to review and rotate. Count them before you sign.

Cleric's list is the widest of the standalone tools: Datadog, Grafana, Prometheus, New Relic, Splunk, Elasticsearch, Honeycomb, Sentry and others on the observability side; AWS, GCP, Kubernetes and Cloudflare; GitHub, Bitbucket, Buildkite and LaunchDarkly; PagerDuty and incident.io; Slack, Teams, Jira, Confluence and Linear; plus a generic MCP connector for internal tools.

Paige pulls logs from Grafana, Datadog, New Relic and CloudWatch, runbooks from Confluence and GitHub, and searches code in connected GitHub repos. One limit buyers miss: it analyses only the first 2,000 characters of each custom_details field and each note, so a long stack trace pasted into an alert payload gets truncated.

Traversal works on top of your existing observability tools and says it needs no agents, sidecars or schemas. Resolve deploys a "satellite" gateway that can read only Kubernetes metadata or proxy data from your on-call tools, and says it does not store raw data. Bits needs the least new plumbing because the data is already in Datadog. Kubernetes one-click actions need a Datadog Private Action Runner with network access to the cluster.

Lock-In: Platform-Native Versus Vendor-Neutral

Platform-native agents are cheaper and easier to start, and they make leaving your observability vendor harder. Bits investigates the telemetry Datadog holds; Paige's memory and playbooks live inside PagerDuty services. If you plan to move even part of your logs to cheaper storage in the next two years, a Datadog-only investigator gets worse every time a data source leaves.

The standalone tools (Cleric, Traversal, Resolve) read whatever backend you run, so the agent survives an observability migration. In the Next Platform survey, 52% said they would consider switching telemetry tools for AI insights that worked across any back end. That is a stated preference rather than a purchase, but it tells you which way the market expects to move.

Our position: if Datadog is your only backend and will stay that way, Bits at about $4 an investigation beats paying a standalone vendor $5 to read the same data through an API. If you run two or more backends, the platform-native discount buys you a blind spot.

Accuracy: Case Studies Versus Your Own Backtest

Every accuracy figure a vendor publishes is self-reported, and none is comparable to another. Cleric says it closed 58% of alerts on its own between May 20 and October 15, 2026, and that 51% of resolved Q3 incidents were fixed with its suggested remediation, both judged by its own resolution monitor. Traversal cites 80% RCA accuracy at PepsiCo and a 38% MTTR cut at DigitalOcean. Datadog reports a 44% higher quality score than its March 2026 release on an internal benchmark. These are useful as a list of what to test. None of them is evidence about your environment.

The failure you are buying insurance against is the confident wrong answer. A practitioner write-up from DevHelm found its agent made claims the metrics did not support in about 5% of sessions, and investigated three simultaneous alerts separately when a human would have seen one upstream cause. Its fix was to require the agent to cite a metric value or log line for every claim.

Run the backtest this way, which follows the historical-replay approach InfoWorld recommends:

  1. Pick 20 to 30 resolved incidents that have a written postmortem with an agreed root cause. Include at least five where the first responder's guess was wrong.
  2. Check that the telemetry still exists. If your log retention is shorter than the incident's age, the agent has nothing to read and the test is void.
  3. Give each vendor the same alert payload and the same read-only scopes, with no hints.
  4. Score three things per incident: correct root cause, evidence that supports it, and time to answer. Count every confident wrong answer as worse than "I don't know."
  5. Record the cost of each run. Resolve shows it live; for credit-based tools, read the meter before and after.

If a vendor will not run on your past incidents during a trial, treat its published accuracy figure as marketing and move on to the next one.

Who Should Not Buy an AI SRE Agent Yet

Skip this purchase if you lack the three inputs these agents run on: runbooks, clean telemetry and an incident history.

Without runbooks, Paige has nothing to execute and the others have no record of what "fixed" looks like. Without consistent service names and trace propagation, the agent reads the same mess your on-call engineer does, faster and with more confidence. Without 20 or so documented incidents, you cannot run the backtest above, which means you would be buying on the demo. Spend that quarter on alert hygiene and postmortems; the agent will score better when you get to it.

Small teams with a handful of incidents a month should also wait. On Cleric's Team plan at $700/month, a team that runs five investigations a month pays $140 per incident.

How to Decide: The Criteria That Predict Regret

The buyers who regret these purchases picked on demo accuracy and skipped the two questions below.

  1. What write access will this need in 12 months, and who approves it? If the answer is "the agent decides," you need Datadog-style guardrails or Cleric-style PRs, and a written permission model from Traversal or Resolve before you sign.
  2. What does the bill do when alert volume triples during a bad week? Cleric's free repeat alerts and Resolve's spend caps handle storms well. Datadog on-demand credits and PagerDuty AI Action packs grow with every request.

What changes the answer: on-prem or bring-your-own-model requirements move you to Traversal or Resolve Custom. A single Datadog backend moves you to Bits. An existing PagerDuty platform contract makes Paige worth turning on for free, though not worth buying a platform tier for.

This Week: Pull 20 resolved incidents with postmortems and confirm the logs and traces for each are still retained.

This Month: Run Cleric and one other agent (Bits if you are on Datadog, Traversal if you need on-prem) against those incidents with read-only scopes, and score root cause, evidence and cost per run.

Before You Sign: Get the write-path permission model in writing, set a monthly spend cap, and model the bill at three times your normal alert volume.

The Bottom Line

AI SRE agents look like the first wave of APM tools did: every vendor claims fast root cause, and the winner for a given team is whichever one fits that team's telemetry and its security review. Cleric is the safest default because you can read its price, its scopes and its fix path before a sales call. Bits is the cheapest if Datadog is your whole world. Traversal and Resolve are for buyers who need deployment control and will negotiate. Paige is worth enabling if you already own the platform tier that includes it.

Score them on your own postmortems before any of them gets write access.

Continue Reading

Share:

Frequently Asked Questions

What is an AI SRE agent?

An AI SRE agent is software that takes a production alert, queries your metrics, logs, traces, deploy history and code, and returns a root-cause hypothesis with evidence. Some, like Cleric and Datadog Bits, can also propose fixes as pull requests or run guardrailed infrastructure actions.

How much does Datadog Bits AI SRE cost?

Bits is billed in Datadog AI Credits: $500 per 500 credits per month on an annual commitment, or $1.30 per credit on-demand, per Datadog's pricing page checked November 19, 2026. An investigation is listed at about 4 credits, and the cost sits on top of your existing Datadog bill.

Which AI SRE agents publish their pricing?

As of November 19, 2026, Cleric ($0.50 per credit, 10 credits per investigation), Datadog (AI Credits), Resolve AI (per-token rates) and PagerDuty (platform tiers with AI Actions) publish prices. Traversal does not and asks buyers to request pricing.

Do AI SRE agents need write access to production?

No. All five agents compared here start read-only. Cleric opens fix PRs only when you allow it, Datadog Bits runs infrastructure actions only inside admin-defined guardrails, and PagerDuty's Paige runs runbook actions with human approval.

How do I test an AI SRE agent's accuracy?

Replay 20 to 30 resolved incidents that have written postmortems, give each vendor the same alert payload and read-only scopes, and score correct root cause, supporting evidence, time to answer and cost per run. Confirm your telemetry for those incidents is still retained first.

Newsletter

More on AI ROI & Cost, every Sunday

The week's enterprise AI news, AI ROI & Cost included, in one Sunday email.

Weekly, free. Unsubscribe any time. Privacy