Anthropic Managed Agents 9-day field test, AI infrastructure on AutoKaam
FIELD NOTE · COVER · MAY 15, 2026 · ISSUE LEAD
FIELD NOTE·May 15, 2026·7 MIN

Anthropic Managed Agents, Nine Days In, The Operator Read

Dreaming and webhooks landed May 6, $0.08 per session hour is the headline, the failure modes are quieter and more interesting.

By·
FIELD NOTEMAY 15, 2026 · AUTOKAAM EDITORIAL

Multiagent Orchestration, Outcomes, and Dreaming, three new features on Managed Agents, announced from the Code with Claude SF keynote, May 6, 2026.

Anthropic

What AutoKaam Thinks
  • Managed Agents launched April 8 at $0.08 per session hour. May 6 added Dreaming, Outcomes, Multiagent Orchestration. Webhooks ship with eight session events, signing secret prefixed whsec_.
  • Operator math: at $0.08 an hour, one agent running 24x7 costs $57.60 a month. Add tokens, web search, and lack of Batch API discount and the bill triples. Claude Code on Max at $200 per month still…
  • Compared to running on Coolify on Oracle ARM with the regular Claude API, hosted only wins when your agent idles more than it runs and webhook ingress matters. Pure batch is cheaper self-hosted.
  • Apply for Dreaming today, ship one webhook-driven workflow this week, do not migrate anything that runs hot.
0.08
USD per hour per agent
ANTHROPIC + INDIAN SOLO FOUNDERS
Named stake

A two-person agency in Bengaluru runs four support agents for a US SaaS client. Each one triages Linear tickets, replies in Slack, drafts the GitHub Issue when needed. Their April bill on the regular Claude API plus a small VPS on Oracle ARM came to ₹38,000. The founder did the math on Managed Agents the morning after the Code with Claude keynote and stayed put. The number that mattered was not $0.08, it was idle time.

This is the nine-day field read. The service went public beta on April 8, 2026. On May 6, Anthropic's San Francisco keynote added three features: Dreaming, Outcomes, Multiagent Orchestration. Webhook ingress for Slack, Linear, GitHub shipped alongside. Pricing has not moved.

What Actually Shipped

Managed Agents is Anthropic's hosted runtime. You define an agent, it runs on their infrastructure, you do not manage the container. Billing is multi-axis: tokens at standard Claude API rates (Opus 4.6 at $5 input, $25 output per million), runtime at $0.08 per session hour billed to the millisecond, tool calls separate. Web search in-session is $10 per 1,000 calls. The 50% Batch API discount does not apply. Bedrock and Vertex routes are not supported.

The May 6 additions:

Dreaming, a scheduled background process that reviews an agent's recent sessions, extracts patterns, and writes plain-text notes plus structured playbooks the next session reads. It does not retrain the model. Harvey, the legal-AI company Anthropic cited on stage, reported task completion up roughly 6x after enabling it. Research preview, gated, you request access.

Outcomes, where you define a rubric, a grader scores the agent's output, and a failure triggers another pass with feedback. Wisedocs reported reviews 50% faster on stage. Research preview.

Multiagent Orchestration, a lead agent fans work to specialist subagents on a shared filesystem. Netflix's platform team is using it for parallel log analysis. Research preview.

Webhooks, the quiet shipper. You register an HTTPS endpoint in the Console, Anthropic mints a signing secret prefixed whsec_, you subscribe to events. The taxonomy: session.status_run_started, session.status_idled, session.status_rescheduled, session.status_terminated, thread variants, session.outcome_evaluation_ended, and vault_credential.refresh_failed. Delivery is at-least-once, retries reuse the same event.id, 5-minute replay window via X-Webhook-Signature. Twenty consecutive failures and the endpoint auto-disables.

The Operator Math

Take the Bengaluru agency. Four agents, each averaging six runtime hours a day across a 12-hour shift (the rest idle, waiting on user replies, tool confirmations, the slow Linear API). At $0.08 per running hour:

  • Runtime per agent: 6 x 30 = 180 hours = $14.40 a month
  • Four agents: $57.60 runtime
  • Tokens (30M input, 4M output per agent per month on Opus 4.6, 80% cache hit): ~$115 per agent, $460 total
  • Web search at 200 a day across the fleet: ~$60
  • Hosted total: ~$577 a month, roughly ₹48,000

Same load on the regular Claude API behind Coolify on Oracle ARM:

  • Tokens: $460
  • Compute: Oracle Always Free ARM tier, ~₹0
  • Engineering tax: one weekend wiring an event queue, one ongoing day a month
  • Self-hosted total: ~$460 plus your time, roughly ₹38,000

Hosted premium is ~₹10,000 a month. Whether it is worth it depends on one question: are you billing the client hourly or by outcome? If outcome, the engineering tax is invisible to revenue and self-hosting wins. If hourly, the saved weekend pays for itself. For a Delhi CA practice running one research agent that pulls case law overnight, idle dominates and there is no in-house engineering, hosted wins easily. For a Mumbai e-commerce ops team batch-processing returns at 2 AM, self-hosted wins because the workload is hot, not idle.

The headline cost lever is not $0.08. It is idle-time accounting. Managed Agents does not bill idle, the regular API has no idle, every call is hot. Spend 70% of wall-clock waiting for a human or tool and hosted is dramatically cheaper than always-on. Run hot 90% of the time and hosted is a tax.

Dreaming, The Contrarian Read

Dreaming is the press feature. The mechanic is a structured note-taking ritual between sessions, not a weight update. Harvey's 6x is the cited evidence. The catch the keynote did not surface: Dreaming is observable, the agent writes plain-text notes and playbooks you can read. Good for audit, less good for moat. A competitor with Outcomes data and one engineer can replicate the playbook layer in a week. The defensibility lives in the data you let the agent see, not the algorithm.

For a small Indian shop, Dreaming earns the request only if your agent touches the same kind of work day after day: support triage seeing the same tickets, a research agent returning to the same companies, a CA's compliance agent walking the same statute each quarter. Novel work every run and Dreaming is theatre.

Webhooks Are The Three-Use-Case Feature

For a 1-3 person Indian operator, three webhook integrations that pay back:

  1. Linear or GitHub Issue triage. New issue fires session.status_run_started, agent triages and labels, session.status_idled fires on completion, you ping the owner in Slack. Replaces the daily standup task nobody loves.
  2. Slack escalation on outcome failure. Outcomes with a rubric, listen for session.outcome_evaluation_ended, on fail ping a human reviewer. The support-quality safety net most Indian SaaS shops bolt on with home-grown cron jobs.
  3. Vault credential rotation alerts. vault_credential.refresh_failed is the unsung event. If you store an OAuth token for a client's CRM and it expires, you find out before the next run, not after.

The trap, called out in the Hookdeck breakdown: the inbound routines.fire endpoint has no idempotency key. Every webhook retry creates a new Claude session. Twenty network blips becomes twenty sessions, twenty runtime charges. Deduplicate on event.id server-side before firing the routine. Biggest footgun in the current beta.

What Anthropic Has Not Shipped

Three gaps worth flagging, all observed, none speculated:

  • No Batch API discount on Managed Agents. If you have a nightly batch workload, you lose the 50% discount the regular API offers. Verified in the WaveSpeed pricing breakdown.
  • No Bedrock or Vertex route. Managed Agents is Claude Platform direct only. Enterprises with AWS or GCP commitments cannot route it through their cloud-credit pool. For Indian startups with leftover GCP credits, this is real money.
  • No idempotency on inbound webhook firings. Documented above. The cookbook recommends deduplicating yourself, which is a polite way of saying it is your problem.

Verdict

Should an Indian solo founder pay for Managed Agents today? One signal: does your agent spend more than half its wall-clock time idle? Yes, hosted is cheaper than equivalent always-on infra and the webhook integration is real leverage. No, stay on the regular Claude API behind Coolify or your own VPS. The $0.08 number is a distraction. The lever is the idle gap.

Dreaming and Outcomes are gated previews. Apply now if you fit the profile, customer-facing agents seeing repetitive work, but do not migrate a hot batch workload because of them. Multiagent Orchestration is interesting if you already have a lead-plus-specialist pattern, otherwise it adds complexity for no immediate gain.

Claude Code on the Max plan at $200 a month is still the right answer for single-seat developer work. Managed Agents is for the workloads where a human is not the loop, the webhook is.

What To Do This Week

  1. Audit your agent fleet for idle ratio. Pull last month's wall-clock vs runtime numbers. Idle over 50% means Managed Agents is worth a pilot.
  2. Wire one webhook-driven workflow on Managed Agents. Linear triage or Slack-escalation-on-outcome-failure. Deduplicate on event.id server-side, this is non-negotiable.
  3. Apply for Dreaming research preview. Two-line application via the Console. If you run a repetitive-work agent and you wait, you go to the back of the queue.

Related